Comsysto Reply GmbH, profile picture
Uploaded byComsysto Reply GmbH
693 views

Data Reliability Challenges with Spark by Henning Kropp (Spark & Hadoop User Group)

The document discusses the challenges and failures of artificial intelligence (AI) regarding trust, ethics, and biases, highlighting instances where AI systems have led to harmful outcomes. It emphasizes the importance of robustness, specification, and assurance to prevent future issues while calling for accountability and ethical considerations in AI development. The conclusion suggests that without trust, AI's advancements may grow indiscriminately, underlining the need for education and awareness on AI's dark sides.

Embed presentation

Download to read offline
1 / 130
2 / 130
3 / 130
4 / 130
5 / 130
6 / 130
7 / 130
8 / 130
9 / 130
10 / 130
11 / 130
12 / 130
13 / 130
14 / 130
15 / 130
16 / 130
17 / 130
18 / 130
19 / 130
20 / 130
21 / 130
22 / 130
23 / 130
24 / 130
25 / 130
26 / 130
27 / 130
28 / 130
29 / 130
30 / 130
31 / 130
32 / 130
33 / 130
34 / 130
35 / 130
36 / 130
37 / 130
38 / 130
39 / 130
40 / 130
41 / 130
42 / 130
43 / 130
44 / 130
45 / 130
46 / 130
47 / 130
48 / 130
49 / 130
50 / 130
51 / 130
52 / 130
53 / 130
54 / 130
55 / 130
56 / 130
57 / 130
58 / 130
59 / 130
60 / 130
61 / 130
62 / 130
63 / 130
64 / 130
65 / 130
66 / 130
67 / 130
68 / 130
69 / 130
70 / 130
71 / 130
72 / 130
73 / 130
74 / 130
75 / 130
76 / 130
77 / 130
78 / 130
79 / 130
80 / 130
81 / 130
82 / 130
83 / 130
84 / 130
85 / 130
86 / 130
87 / 130
88 / 130
89 / 130
90 / 130
91 / 130
92 / 130
93 / 130
94 / 130
95 / 130
96 / 130
97 / 130
98 / 130
99 / 130
100 / 130
101 / 130
102 / 130
103 / 130
104 / 130
105 / 130
106 / 130
107 / 130
108 / 130
109 / 130
110 / 130
111 / 130
112 / 130
113 / 130
114 / 130
115 / 130
116 / 130
117 / 130
118 / 130
119 / 130
120 / 130
121 / 130
122 / 130
123 / 130
124 / 130
125 / 130
126 / 130
127 / 130
128 / 130
129 / 130
130 / 130
The Dark Sides of AI Where AI went wrong and how to handle it. Alexander Pospiech ,
Data Engineer/Scientist @ inovex Security and Privacy Apologist Ɣ Father of One Dinghy-Sailor  Nerd 2
What is trust? 1 1 What will it take for us to trust AI? by Alan Finkel3
trust noun the belief that you can trust someone or something trust verb to believe that someone is good and honest and will not harm you, or that something is safe and reliable 2 2 https://dictionary.cambridge.org/dictionary/english/trust4
People are comfortable with algorithmic decisions, sometimes more than human advice. 3 3 Do People Trust Algorithms More Than Companies Realize?5
How it already has gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 6
How it already has gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 7
Sugar bananas by Maksym Kozlenko licensed as CC-BY-SA-4.0
The Cautious Path to Strate- gic Advantage: How Militaries Should Plan for AI by Peter Eckersley
Chinese facial recognition system confuses bus ad with a jaywalker by Jon Fingas
Turkish is a gender neutral language by Alex Shams (28.11.2017) 11
Google is fixing gender bias in its Translate service by Ivan Mehta 12
Google’s technology will not suggest gender-based pronouns because the risk is too high that its “Smart Compose” technology might predict someone’s sex or gender identity incorrectly and offend users, product leaders ... 4 4 Fearful of bias, Google blocks gender-based pronouns from new AI tool by Paresh Dave13
The Machine Fired Me - No human could do a thing about it! Necessary orders are sent automatically and each order completion triggers another order. 4 idiallo.com/blog/when-a-machine-fired-me by Ibrahim Diallo14
Responses to Critiques on Machine Learning of Criminality Perceptions by Xiaolin Wu, Xi Zhang 15
How it already has gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 16
Racing Stripes Car Top View by qubodup original art: Autonomous Trap 001 by James Bridle (2017) 17
The Moral Machine experiment by Edmond Awad, Sohan Dsouza, Richard Kim, Jonathan Schulz, Joseph Henrich, Azim Shariff, Jean-François Bonnefon & Iyad Rahwan18
https://twitter.com/TayandYou (2016)
Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images. by Nguyen A, Yosinski J, Clune J. at Evolving AI Lab, University of Wyoming 20
Explaining and harnessing adversarial examples. by Goodfellow, Ian J., Jonathon Shlens, and Christian Szegedy. at OpenAI 21
Strike (with) a Pose: Neural Networks Are Easily Fooled by Strange Poses of Familiar Objects by Michael A. Alcorn, Qi Li, Zhitao Gong, Chengfei Wang, Long Mai, Wei-Shinn Ku, Anh Nguyen 22
Fooling Neural Networks in the Physi- cal World with 3D Adversarial Objects by Anish Athalye, Logan Engstrom, Andrew Ilyas & Kevin Kwok at LabSix
Adversarial Reprogramming of Neural Networks. by Gamaleldin F. Elsayed, Ian Goodfellow and Jascha Sohl-Dickstein at Google Brain 24
Robust Physical-World Attacks on Deep Learning Models by Kevin Eykholt, Ivan Evtimov, Ear- lence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, Dawn Song
Accessorize to a Crime: Real and Stealthy Attacks on State-of-the-Art Face Recognition by Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, Michael K. Reiter
Juggalo by twitter.com/tahkion27
Malware: DeepLocker - Concealing Targeted Attacks with AI Locksmithing DEF CON 25 - Hyrum Anderson - Evading next gen AV using AI Intrusion Detection: Here’s The Truth, GANs Easily Can Fool Intrusion Detection Systems Anti-Virus: G Data uses nerual network based DeepRay for detecting malware 5 5 Antiviren-Software mit KI-Technik: G Data setzt auf DeepRay28
Fake fingerprints can imitate real ones in biometric systems – research by Philip Bontrager, Aditi Roy, Julian Togelius, Nasir Memon, Arun Ross
Giphy30
America’s biggest body-camera company says facial recognition isn’t accurate enough for policing decisions 5 America’s biggest body-camera company says facial recognition isn’t accurate enough for policing decisions by Dave Gershgorn31
Face Recognition Field Test at Südkreuz 12 by C.Suthorn under CC-BY-SA-4.0
While the authors applaud the force for attempting to develop an ethically sound and legally compliant approach to predictive policing, they warn that the ethical principles in the proposal are not developed enough to deal with the broad challenges this kind of technology could throw up, and that “frequently the details are insufficiently fleshed out and important issues are not fully recognized.” 6 6 Britain Is Developing an AI-Powered Predictive Policing System33
How it already has gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 34
Gender Shades by Joy Buolamwini (2018) and her MIT group
Google Photos, y’all fucked up by Jacky Alcine (28.06.2015) 36
Amazon’s Disturbing Plan to Add Face Surveillance to Your Front Door37
Amazon’s Face Recognition Falsely Matched 28 Members of Congress With Mugshots 38
Amazon’s Face Recognition Falsely Matched 28 Members of Congress With Mugshots39
I tested 14 sentences for "perceived toxicity" using Perspectives. jessamyn west (25.08.2017) 40
Design Justice, A.I., and Escape from the Matrix of Domination by Sasha Costanza-Chock 41
Amazon.com Inc’s machine-learning specialists uncovered a big problem: their new recruiting engine did not like women. 6 Reuters - RPT-INSIGHT-Amazon scraps secret AI recruiting tool that showed bias against women42
Predictive Policing 43
... predictive models reinforce existing police practices ... ... patterns of police records, not patterns of crime ... ... cannot predict patterns of crime that are different ... 7 7 USA by Human Rights Data Analysis Group44
... the differences in arrest rates by ethnic group between predictive policing and standard patrol practices were not statistically significant, ..." 8 8 Field-data Study Finds No Evidence of Racial Bias in Predictive Policing (2018) by Forensic Magazine45
Predictice Judgement 46
Positive Positive: equal chance False Positives/Negatives: unfair 9 9 How to Fight Bias with Predictive Policing (2018) by Eric Siegel in Scientific American47
... COMPAS is no more accurate or fair than predictions made by people with little or no criminal justice expertise. 10 10 The accuracy, fairness, and limits of predicting recidivism (2018) by Julia Dressel and Hany Farid in Science Advances48
... despite COMPAS’s collection of 137 features, the same accuracy can be achieved with a simple linear classifier with only two features.11 11 The accuracy, fairness, and limits of predicting recidivism (2018) by Julia Dressel and Hany Farid in Science Advances49
The “bias” comes from base rates. 12 12 AI Ethics, Impossibility Theorems and Tradeoffs50
Faceception ...recognizing “High IQ”, “White-Collar Offender”, “Pedophile”, and “Terrorist” ... According to Social and Life Science research personalities are affected by genes. Our face is a reflection of our DNA. 13 13 Faception51
Predictive Staatsschutz GES-3D - Multi-Biometrische Gesichtserkennung MisPel-Projekt - Multi-Biometrisierte Forensische Personensuche in Lichtbild- und Videomassendaten INTEGER - Visuelle Entscheidungsunterstützung bei der Auswertung von Daten aus sozialen Netzwerken PANDORA - Propaganda, Mobilisierung und Radikalisierung zur Gewalt in der virtuellen und realen Welt X-SONAR - Extremistische Bestrebungen in Social Media Netzwerken RADIG-Z -Radikalisierung im digitalen Zeitalter RISKANT - Risikoanalyse bei islamistisch motivierten Tatgeneigten Survey: Predictive Staatsschutz auf Telepolis von Matthias Becker 52
How it already has gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 53
Everbody has a deep learning paper Thousands of papers ( 20.000/year) big part have low quality faked results or just lack of knowledge 54
Agenda How it already has gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion55
How it already has gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 56
Quadrants Intended Unintended Inside Black Hat AI Bias a.k.a Vendor/User killer robots bad validation Cheating14 wrong use of model Outside Adversarial Attacks Bias a.k.a Attacker/User data poisoning 14 US border agents hacked their “risk assessment” system to recommend detention 100% of the time57
Roles & Accountability Researchers Developers Vendors Operators Users Regulators Attacker 58
Accountability - A scenario build an AI for fraud detection for a bank AI blocks tranaction of company A A goes bankrupt A sues the bank developer (you) is questioned as a witness Can you provide answers? 15 15 Hype-Tech by Felix von Leitner (translated from german)59
Accountability - and how? reporting process? transparency to the public? power to decide on necessary changes? provability of work afterwards? 60
Cost of Misbehaving AI Why should the vendor change something? legal consequences ethical/morally behaviour loss of reputation loss of opportunities loss of money 61
Cost of Misbehaving AI Why should the users demand change? loss of money loss of security loss of freedom 62
Defining and Researching "The Game" Defending is always harder than attacking 16 AI for/against: Detection Prediction Prevention Response 16 Is attacking machine learning easier than defending it?63
Regulation - GDPR "Right to be forgotten/"Right to erasure" Älgorithmic Fairnessänd "The Right to Explanation" 64
Regulation - Understanding White House report: Preparing for the future of Artificial Intelligence House of Lords report: AI in the UK: ready, willing and able? Bundestag then: some talk and a list of experts Bundestag now: KI-Eckpunktepapier 65
How it already has gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 66
Specification Design - bugs, ambiguities, side effects, high level language, preference learning, design protocols Emergent - wireheading, delusions, meta-Learning and sub-agents, detecting emerging behavior, reward hacking 16 Building safe artificial intelligence: specification, robustness, and assurance by DeepMind Safety Research - Pedro A. Ortega, Vishal Maini, and the DeepMind safety team67
How it already has gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 68
Robustness Risk and prevention - limitation, risk sensitivity, uncertainty estimates, safety margins, safe exploration, adversaries, cautious generalisation, verification Recovery and stability - availability, instability, error correction, fail-safe, distributional shift, graceful degradation 16 Building safe artificial intelligence: specification, robustness, and assurance by DeepMind Safety Research - Pedro A. Ortega, Vishal Maini, and the DeepMind safety team69
Robustness - Adversaries image, video, malware, health records, ... Find, investigate, train on and robustify against attack vectors. defenses tend to be broken easily learn the adversaries as counter examples use regulization cryptography 17 17 Defense against adversarial attacks using machine learning and cryptography by Ingrid Fadelli70
Robustness - Availability Availability of the processing? Can I DOS a Neural Network? Availability of predcitions or decisions? Is a poisoning attack a DOS? 71
How it already has gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 72
Ethics - discussional approach include other groups than engineers use checklists, like deon - An ethics checklist for data scientists there is no, I repeat NO, programmable ethics! ethics is a discussion and by this an ongoing process 73
Ethics - technical approach enforce consent, clarity, consistency, control, consequences 18 18 Care about AI ethics? What you can do, starting today by Steven Adler74
Chris Anderson: “with enough data, the numbers speak for themselves.” 18 The Hidden Biases in Big Data by Kate Crawford75
Kate Crawford: ẞadly, they can’t. Data and data sets are not objective; they are creations of human design." 18 The Hidden Biases in Big Data by Kate Crawford76
Fairness - Definition Anti-classification Classification parity Calibration: If an algorithm produces a “score,” that “score” should mean the same thing for different groups. 77
Fairness - Tools Aequitas - open source bias audit toolkit IBM - AI Fairness 360 Facebook creates internal tool - Fairness Flow "Microsoft is creating an oracle for catching biased AI algorithms" Tool: What-If tool deon - An ethics checklist for data scientists Fix bias in AI with more tech (probably AI) sigh 78
How it already has gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 79
Assurance Monitoring - interpretability, behavioural screening, traces, estimates of causal influence, tripwires & honeypots Enforcement - testability, interuptibility, boxing, physical security, encryption, signatures, authorisation, human override 18 Building safe artificial intelligence: specification, robustness, and assurance by DeepMind Safety Research - Pedro A. Ortega, Vishal Maini, and the DeepMind safety team80
Assurance - Testing why not currently? what to build? what to test? continuous regression tests? special items? biases? 81
Assurance - Physical Security A neural network is some files on hardware. Can be copied, stolen, modified, ... 82
Assurance - Interpretability The Mythos of Model Interpretability 19 expect superhuman standards, we only tend to have (post-rationalization) but needed for Debugging, Verification, Fairness, Transparency 19 The Mythos of Model Interpretability by Zachary C. Lipton83
Assurance - Interpretability intrinsic vs. posthoc outcome: statistics, visualization, weights, data points, intrinsical model model-specific vs. model agnostic local vs. global remind the the needs and the knowledge of the different roles 84
Introduction to Local Interpretable Model-Agnostic Explanations (LIME) (2016) by Marco Tulio Ribeiro, Sameer Singh, Carlos Guestrin in O’Reilly
Introduction to Local Interpretable Model-Agnostic Explanations (LIME) (2016) by Marco Tulio Ribeiro, Sameer Singh, Carlos Guestrin in O’Reilly
Question: I have an unexplainable AI that detects 1000 tumors and an explainable one that detects 100. Which one do I prefer? 87
How it already has gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 88
Development ETL - biases, fairness, privacy, data poisoning, ... Reproducibility - model management, data management, result management, ... 89
In many real-world cases, the researcher won’t have made notes or remember exactly what she did, so even she won’t be able to reproduce the model. 20 20 The Machine Learning Reproducibility Crisis (2018) by Pete Warden90
Yet AI researchers say the incentives are still not aligned with reproducibility. 21 21 Missing data hinder replication of artificial intelligence studies (2018) by Matthew Hutson in Science91
Privacy - Confidentiality Privacy Encryption Possible tools: Differential Privacy Homomorphic Encryption Federated learning 22 22 Federated Learning: Collaborative Machine Learning without Centralized Training Data92
How it already has gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 93
I would say: It’s not the TÜV! 94
KI Bundesverband? 95
How it already has gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 96
Education Educate AI basics as part of data literacy in school and college. 97
What can you do? Educate Warn Support Research Develop 98
Without our trust AI will grow regardlessly. 99
With the stated advancements AI will have our trust and may work like expected. 100
Thank you Alexander Pospiech Data Engineer inovex GmbH Lindberghstraß e 3 80939 München alexander.pospiech@inovex.de +49 153 - 31 81 051
Conferences and Meetings I Specific on the Dark Sides: Conference on Fairness, Accountability, and Transparency FATML - Fairness, Accountability, and Transparency in Machine Learning Interpretable ML Symposium @NIPS NIPS 2017 Tutorial - Fairness in Machine Learning Reproducibility in ML Workshop, ICML’18 IEEE 1st Deep Learning and Security Workshop Data Ethics workshop, KDD 2014 MAKE-Explainable AI Advances on Explainable Artificial Intelligence 102
Conferences and Meetings II Nemesis ’18 - 1st Workshop on Recent Advances in Adversarial Machine Learning NIPS 2018 Workshop on Security in Machine Learning Generic on AI: AI for Good Global Summit 103
Conferences and Meetings I General on Security: CCC DefCon SHA BlackHat 104
Research Groups and Organizations I AI specific: AINow - A research institute examining the social implications of artificial intelligence Evolving AI Lab, University of Wyoming OpenAI LabSix EFF on Artificial Intelligence & Machine Learning EFF - AI Progress Measurement EvalAI - Evaluating state of the art in AI EvadeML - Machine Learning in the Presence of Adversaries Adversarial Machine Learning, Università degli Studi di Cagliari 105
Research Groups and Organizations II SunBlaze at UCB Diskriminierung durch KI (Künstliche Intelligenz) (DiKI) Algorithmische Gegenmacht Center for Human-Compatible AI Technische und rechtliche Betrachtungen algorithmischer Entscheidungsverfahren by GI Algorithmenethik 106
Research Groups and Organizations I General: Trusted AI - IBM Research is building and enabling AI solutions people can trust Ethics/Fairness: SIENNA: Technology, ethics and human rights Human Rights Data Analysis Group AlgorithmWatch Netzpolitik on Predictive Policing 107
Projects I Interpretability: Project: DARPAs Explainable Artificial Intelligence Project: Explain AI Project: DALEX: Descriptive mAchine Learning EXplanations 108
Communities I OpenMined 109
Classes I CS 294: Fairness in Machine Learning, UC Berkeley 18739 Security and Fairness of Deep Learning, Carnegie Mellon Adversarial and Secure Machine Learning IEEE’s Artificial Intelligence and Ethics in Design The Ethics and Governance of Artificial Intelligence Attacking Networks with Adversarial Examples CS 294-149: Safety and Control for Artificial General Intelligence (Fall 2018) Machine-learning - crash-course -fairness 110
Themensammlung I Netzpolitik on Predictive Policing EFF on Artificial Intelligence & Machine Learning EFF - AI Progress Measurement EvalAI - Evaluating state of the art in AI AI safety resources Ethics: AI Ethics Resources 111
Github with Lists I General: Awful AI Ethics: Data Science and Ethics Resources Global Data Ethics Pledge (GDEP) ML and Security Machine Learning for Cyber Security Awesome Awesome AI Security Adversarial Attacks: Awesome Adversarial Examples for Deep Learning Awesome Adversarial Machine Learning 112
Github with Lists II Awesome Machine Learning for Cyber Security Privacy; awesome-ai-privacy Private-Ai-Resources Machine Learning Ethics References Fairness: Toward ethical, transparent and fair AI/ML: a critical reading list for engineers, designers, and policy makers Interpretability; Awesome Interpretable Machine Learning Awesome awesome-machine-learning-interpretability Awesome 113
Github with Code I Ethics: deon - An ethics checklist for data scientists An Open Standard for Ethical Enterprise-Grade AI Interpretability: H20.ai: Machine Learning Interpretability (MLI) Explanation Explorer Interpretable Machine Learning with Python iml: interpretable machine learning ML Insights Fairness: Comparing fairness-aware machine learning techniques. 114
Github with Code II Themis ML - Fairness-aware Machine Learning Adversarial Attacks: Introduction to Adversarial Machine Learning DeepFool CleverHans EvadeML-Zoo AdvFlow Evaluating and Understanding the Robustness of Adversarial Logit Pairing IBM adversarial-robustness-toolbox deep-pwning FoolBox, 115
Papers I LIME, Local Interpretable Model-Agnostic Explanations (2016) 116
Studies I European expert group seeks feedback on draft ethics guidelines for trustworthy artificial intelligence 117
Blogs I a blog about security and privacy in machine learning MLSec covert.io security + big data + machine learning Data Driven Security Automating OSINT BigSnarf Security of Machine Learning 118
Slides I General: Hype-Tech by Felix von Leitner Adversarial Attacks: How CLEVER is your neural network? Robustness evaluation against adversarial examples by Pin-Yu Chen at IBM Research AI Privacy: Machine Learning & Privacy: It’s Complicated Fairness: AI Ethics, Impossibility Theorems and Tradeoffs 119
Blog Posts I General: 6 core falsehoods about the digital sphere Nein, Ethik kann man nicht programmieren You created a machine learning application. Now make sure it’s secure. Tooling: Deep automation in machine learning Fainess: Help Wanted: An Examination of Hiring Algorithms, Equity, and Bias (December 2018) Security: Attacks against machine learning — an overview 120
Blog Posts II Machine Learning for Cybersecurity 101 Interpretability: Testing machine learning explanation techniques Explainable AI and the Legality of Autonomous Weapon Systems Adversarial Attacks: The Definitive Security Data Science and Machine Learning Guide Discussion and Survey of Adversarial Examples and Robustness in Deep Learning A Brief Introduction to Adversarial Examples Adversarial Sample Generation: Making Machine Learning Systems Robust for Security Survey on Security and Privacy of Machine Learning Attacking Machine Learning Detectors: the state of the art review121
Podcasts I Fairness: O’Reilly Data Show Podcast - Why it’s hard to design fair machine learning models Privacy: How privacy-preserving techniques can lead to more robust machine learning models 122
Videos - general I General: Youtube: Stephen Fry describing our future with artificial intelligence and robots 34c3 - Beeinflussung durch Künstliche Intelligenz 34c3 - Deep Learning Blindspots SHA2017 - The Security and Privacy Implications of AI and Machine Learning Youtube - DEF CON 24 - Clarence Chio - Machine Duping 101: Pwning Deep Learning Systems Youtube: Do You Trust This Computer? TED - The era of blind faith in big data must end Managing Risk in Machine Learning - Ben Lorica (O’Reilly Media) 123
Videos - general II Ethics: Ethik der Algorithmen - Tom Hillenbrand zum "Schreckensszenario KI" Interpretability: [HUML16] 06: Zackary C. Lipton, The mythos of model interpretability "Why Should I Trust you?" Explaining the Predictions of Any Classifier, KDD 2016 Interpretable Machine Learning Using LIME Framework - Kasia Kulma (PhD), Data Scientist, Aviva TWiML - Trust in Prediction of Machine Learning Models - EMEA Meetup #4 - December 2018 Explaining Complex Machine Learning Models With LIME - Shirin Glander (codecentric AG) 124
Videos - general III Fairness: Tutorial: 21 fairness definitions and their politics Adversarial Examples: Is "Adversarial Examples" an Adversarial Example? How to Successfully Harness AI to Combat Fraud and Abuse - RSA 2018 Adversarial Robustness, Theory and Practice Accountability: Lack of Accountability in Data Science: Why We Should All Be Building Inclusive AI 125
Adversarial Examples - Examples I Audio Adversarial Examples 126
Adversarial Attack Competitions I MNIST Adversarial Examples Challenge NIPS 2017 Competition: Non-targeted Adversarial Attack Introducing the Unrestricted Adversarial Examples Challenge 127
Lists I Reward Hacking Specification gaming examples in AI 128
News I Autonomous Driving Uber self-driving car kills a pedestrian Incidents with Tesla Autopilot Even the cute Waymo 129
Online Books I Interpretable Machine Learning Fairness and machine learning 130

More Related Content

PPTX
Responsible AI in Industry (Tutorials at AAAI 2021, FAccT 2021, and WWW 2021)
byKrishnaram Kenthapadi
 
PDF
AIF360 - Trusted and Fair AI
byAnimesh Singh
 
PDF
How do we train AI to be Ethical and Unbiased?
byMark Borg
 
PPTX
Transparency in ML and AI (humble views from a concerned academic)
byPaolo Missier
 
PDF
Interpretability beyond feature attribution quantitative testing with concept...
byMLconf
 
PPTX
BY GROUP-9_2025091yuggy0050209_0000.pptx
bykorad0417
 
PPTX
CS Org Ldrshp Poster
byTurner Sparks
 
PDF
Stay Safe and Healthy with Computer Vision
byNUS-ISS
 
Responsible AI in Industry (Tutorials at AAAI 2021, FAccT 2021, and WWW 2021)
byKrishnaram Kenthapadi
 
AIF360 - Trusted and Fair AI
byAnimesh Singh
 
How do we train AI to be Ethical and Unbiased?
byMark Borg
 
Transparency in ML and AI (humble views from a concerned academic)
byPaolo Missier
 
Interpretability beyond feature attribution quantitative testing with concept...
byMLconf
 
BY GROUP-9_2025091yuggy0050209_0000.pptx
bykorad0417
 
CS Org Ldrshp Poster
byTurner Sparks
 
Stay Safe and Healthy with Computer Vision
byNUS-ISS
 

Similar to Data Reliability Challenges with Spark by Henning Kropp (Spark & Hadoop User Group)

PPTX
Responsible AI
byNeo4j
 
PPTX
Responsible AI in Industry (ICML 2021 Tutorial)
byKrishnaram Kenthapadi
 
PPTX
Adversarial Attacks and Defense
byKishor Datta Gupta
 
PPTX
Tecnologías emergentes: priorizando al ciudadano
byComisión de Regulación de Comunicaciones
 
PDF
Ethical-considerations-in-AIshjsjsbdh.pdf
byRuthPeralta22
 
PDF
Tru_Shiralkar_Gen AI Sec_ ISACA 2024.pdf
byTrupti Shiralkar, CISSP
 
PDF
How to undermine AI dystopia
byBianca Ximenes
 
PDF
When the AIs failures send us back to our own societal biases
byClément DUFFAU
 
PPTX
3 Steps To Tackle The Problem Of Bias In Artificial Intelligence
byBernard Marr
 
PDF
AI NOW REPORT 2018
byPeerasak C.
 
PDF
AI/Data Analytics (AIDA): Key concepts, examples & risks
bySimon Buckingham Shum
 
PDF
“Identifying and Mitigating Bias in AI,” a Presentation from Intel
byEdge AI and Vision Alliance
 
PDF
"Towards Value-Centric Big Data" e-SIDES Workshop - "The dangers of tech-dete...
bye-SIDES.eu
 
PDF
Ethical Dilemmas in AI/ML-based systems
byDr. Kim (Kyllesbech Larsen)
 
PDF
20240705 QFM024 Irresponsible AI Reading List June 2024
byMatthew Sinclair
 
PDF
Evolution of Bias in Artificial Intelligence and Solution the big problem
byAbdulla Hisham
 
PPTX
Ameaças e oportunidades das novas tecnologias para o desenvolvimento e a demo...
byFundação Fernando Henrique Cardoso
 
DOCX
Diginomica 2019 2020 ai ai ethics neil raden articles links and captions
byNeil Raden
 
PDF
Ai Now institute 2017 report
byWilly Marroquin (WillyDevNET)
 
PPTX
Every thing about Artificial Intelligence
byVaibhav Mishra
 
Responsible AI
byNeo4j
 
Responsible AI in Industry (ICML 2021 Tutorial)
byKrishnaram Kenthapadi
 
Adversarial Attacks and Defense
byKishor Datta Gupta
 
Tecnologías emergentes: priorizando al ciudadano
byComisión de Regulación de Comunicaciones
 
Ethical-considerations-in-AIshjsjsbdh.pdf
byRuthPeralta22
 
Tru_Shiralkar_Gen AI Sec_ ISACA 2024.pdf
byTrupti Shiralkar, CISSP
 
How to undermine AI dystopia
byBianca Ximenes
 
When the AIs failures send us back to our own societal biases
byClément DUFFAU
 
3 Steps To Tackle The Problem Of Bias In Artificial Intelligence
byBernard Marr
 
AI NOW REPORT 2018
byPeerasak C.
 
AI/Data Analytics (AIDA): Key concepts, examples & risks
bySimon Buckingham Shum
 
“Identifying and Mitigating Bias in AI,” a Presentation from Intel
byEdge AI and Vision Alliance
 
"Towards Value-Centric Big Data" e-SIDES Workshop - "The dangers of tech-dete...
bye-SIDES.eu
 
Ethical Dilemmas in AI/ML-based systems
byDr. Kim (Kyllesbech Larsen)
 
20240705 QFM024 Irresponsible AI Reading List June 2024
byMatthew Sinclair
 
Evolution of Bias in Artificial Intelligence and Solution the big problem
byAbdulla Hisham
 
Ameaças e oportunidades das novas tecnologias para o desenvolvimento e a demo...
byFundação Fernando Henrique Cardoso
 
Diginomica 2019 2020 ai ai ethics neil raden articles links and captions
byNeil Raden
 
Ai Now institute 2017 report
byWilly Marroquin (WillyDevNET)
 
Every thing about Artificial Intelligence
byVaibhav Mishra
 

More from Comsysto Reply GmbH

PDF
Data lake vs Data Warehouse: Hybrid Architectures
byComsysto Reply GmbH
 
PDF
Software Architecture and Architectors: useless VS valuable
byComsysto Reply GmbH
 
PDF
Distributed Computing and Caching in the Cloud: Hazelcast and Microsoft
byComsysto Reply GmbH
 
PPTX
Ein Prozess lernt laufen: LEGO Mindstorms Steuerung mit BPMN
byComsysto Reply GmbH
 
PDF
ljug-meetup-2023-03-hexagonal-architecture.pdf
byComsysto Reply GmbH
 
PPTX
Apache Apex: Stream Processing Architecture and Applications
byComsysto Reply GmbH
 
PDF
Invited-Talk_PredAnalytics_München (2).pdf
byComsysto Reply GmbH
 
PPTX
Java 9 Modularity and Project Jigsaw
byComsysto Reply GmbH
 
PDF
Alles offen = gut(ai)
byComsysto Reply GmbH
 
PDF
Java cro 2016 - From.... to Scrum by Jurica Krizanic
byComsysto Reply GmbH
 
PDF
Architectural Decisions: Smoothly and Consistently
byComsysto Reply GmbH
 
PDF
21.04.2016 Meetup: Spark vs. Flink
byComsysto Reply GmbH
 
PDF
Bable on Smart City Munich Meetup: How cities are leveraging innovative partn...
byComsysto Reply GmbH
 
PDF
Grundlegende Konzepte von Elm, React und AngularDart 2 im Vergleich
byComsysto Reply GmbH
 
PDF
MicroFrontends für Microservices
byComsysto Reply GmbH
 
PDF
Building a fully-automated Fast Data Platform
byComsysto Reply GmbH
 
PDF
Geospatial applications created using java script(and nosql)
byComsysto Reply GmbH
 
PDF
Smart City Munich Kickoff Meetup
byComsysto Reply GmbH
 
PDF
"Hadoop Data Lake vs classical Data Warehouse: How to utilize best of both wo...
byComsysto Reply GmbH
 
PDF
Spark RDD-DF-SQL-DS-Spark Hadoop User Group Munich Meetup 2016
byComsysto Reply GmbH
 
Data lake vs Data Warehouse: Hybrid Architectures
byComsysto Reply GmbH
 
Software Architecture and Architectors: useless VS valuable
byComsysto Reply GmbH
 
Distributed Computing and Caching in the Cloud: Hazelcast and Microsoft
byComsysto Reply GmbH
 
Ein Prozess lernt laufen: LEGO Mindstorms Steuerung mit BPMN
byComsysto Reply GmbH
 
ljug-meetup-2023-03-hexagonal-architecture.pdf
byComsysto Reply GmbH
 
Apache Apex: Stream Processing Architecture and Applications
byComsysto Reply GmbH
 
Invited-Talk_PredAnalytics_München (2).pdf
byComsysto Reply GmbH
 
Java 9 Modularity and Project Jigsaw
byComsysto Reply GmbH
 
Alles offen = gut(ai)
byComsysto Reply GmbH
 
Java cro 2016 - From.... to Scrum by Jurica Krizanic
byComsysto Reply GmbH
 
Architectural Decisions: Smoothly and Consistently
byComsysto Reply GmbH
 
21.04.2016 Meetup: Spark vs. Flink
byComsysto Reply GmbH
 
Bable on Smart City Munich Meetup: How cities are leveraging innovative partn...
byComsysto Reply GmbH
 
Grundlegende Konzepte von Elm, React und AngularDart 2 im Vergleich
byComsysto Reply GmbH
 
MicroFrontends für Microservices
byComsysto Reply GmbH
 
Building a fully-automated Fast Data Platform
byComsysto Reply GmbH
 
Geospatial applications created using java script(and nosql)
byComsysto Reply GmbH
 
Smart City Munich Kickoff Meetup
byComsysto Reply GmbH
 
"Hadoop Data Lake vs classical Data Warehouse: How to utilize best of both wo...
byComsysto Reply GmbH
 
Spark RDD-DF-SQL-DS-Spark Hadoop User Group Munich Meetup 2016
byComsysto Reply GmbH
 

Recently uploaded

PPTX
WIELD OF STATISTICS IN DECISION - MAKING
bysaberunnisaa
 
PDF
Document 2.pdf bizimkiler yapmış yine dvfweefsdc
byAlperturkoglu
 
PDF
Advanced Data Automation, Orchestration & Cloud Analytics | Wisecor Transform...
bywisecortransformatio
 
PPTX
Ritik Jangid Portfolio, a business analyst
byjangidritik851
 
PDF
What the Insurance Industry Considers “High-Risk” Dog Breeds
byBisnar Chase Personal Injury Attorneys
 
PDF
[Redis Released]- FalkorDB - Redis + Graph Agentic Memory’s Secret Sauce
byGuy Korland
 
PPTX
statistics importance and scope of stats
byfizza2297
 
PDF
UG Syllabus 2023-24 UG Syllabus 2023-24UG Syllabus 2023-24
byECENVSatyanarayanaMu
 
PDF
UG ECE SYLLABUS 2022-26 batchUG ECE SYLLABUS 2022-26 batch
byECENVSatyanarayanaMu
 
PPTX
Health Insight Survey: an introduction | Accessible slides
byOffice for National Statistics
 
PPTX
Electricity Stats by Generation Method - August 2025 - OECD
byPaulYoung221210
 
PPTX
Basics of Research Methodology Part1.pptx
byDepartment of Agricultural Marketing
 
DOCX
THESIS ABSTRACT ABOUT CHALLENGES ENCOUNTERED BY POLICE OFFICERS.docx
byssuser6a111b
 
PPTX
Data_Analysis_Plan-corrected-Biostatistics.pptx
byFahmida Swati
 
PPTX
Correlation-Regression analysis -16.11.25.pptx
byFahmida Swati
 
PPT
PETROLEUM TRAPS.pptPETROLEUM TRAPS.pptmb
byalialialsatre
 
PPTX
Capstone_Presentation.pptxzzzzzzzzzzzzzz
byharbingernoobsaibot
 
PPTX
TERMA DATIS PPT IN AVIATION SECTOR .pptx
byjimmylal
 
PPT
monocot-vs-dicot.ppt FOR I'M IRRIGATION I LOVE THIS SO BAD BUT THIS ONE WAS
bydaaboulaziz5
 
PPTX
MASS INTERPRETATION OF ORGANIC COMPOUNDS
byDevipriya787793
 
WIELD OF STATISTICS IN DECISION - MAKING
bysaberunnisaa
 
Document 2.pdf bizimkiler yapmış yine dvfweefsdc
byAlperturkoglu
 
Advanced Data Automation, Orchestration & Cloud Analytics | Wisecor Transform...
bywisecortransformatio
 
Ritik Jangid Portfolio, a business analyst
byjangidritik851
 
What the Insurance Industry Considers “High-Risk” Dog Breeds
byBisnar Chase Personal Injury Attorneys
 
[Redis Released]- FalkorDB - Redis + Graph Agentic Memory’s Secret Sauce
byGuy Korland
 
statistics importance and scope of stats
byfizza2297
 
UG Syllabus 2023-24 UG Syllabus 2023-24UG Syllabus 2023-24
byECENVSatyanarayanaMu
 
UG ECE SYLLABUS 2022-26 batchUG ECE SYLLABUS 2022-26 batch
byECENVSatyanarayanaMu
 
Health Insight Survey: an introduction | Accessible slides
byOffice for National Statistics
 
Electricity Stats by Generation Method - August 2025 - OECD
byPaulYoung221210
 
Basics of Research Methodology Part1.pptx
byDepartment of Agricultural Marketing
 
THESIS ABSTRACT ABOUT CHALLENGES ENCOUNTERED BY POLICE OFFICERS.docx
byssuser6a111b
 
Data_Analysis_Plan-corrected-Biostatistics.pptx
byFahmida Swati
 
Correlation-Regression analysis -16.11.25.pptx
byFahmida Swati
 
PETROLEUM TRAPS.pptPETROLEUM TRAPS.pptmb
byalialialsatre
 
Capstone_Presentation.pptxzzzzzzzzzzzzzz
byharbingernoobsaibot
 
TERMA DATIS PPT IN AVIATION SECTOR .pptx
byjimmylal
 
monocot-vs-dicot.ppt FOR I'M IRRIGATION I LOVE THIS SO BAD BUT THIS ONE WAS
bydaaboulaziz5
 
MASS INTERPRETATION OF ORGANIC COMPOUNDS
byDevipriya787793
 

Data Reliability Challenges with Spark by Henning Kropp (Spark & Hadoop User Group)

  • 1.
    The Dark Sidesof AI Where AI went wrong and how to handle it. Alexander Pospiech ,
  • 2.
    Data Engineer/Scientist @inovex Security and Privacy Apologist Ɣ Father of One Dinghy-Sailor  Nerd 2
  • 3.
    What is trust?1 1 What will it take for us to trust AI? by Alan Finkel3
  • 4.
    trust noun the belief thatyou can trust someone or something trust verb to believe that someone is good and honest and will not harm you, or that something is safe and reliable 2 2 https://dictionary.cambridge.org/dictionary/english/trust4
  • 5.
    People are comfortablewith algorithmic decisions, sometimes more than human advice. 3 3 Do People Trust Algorithms More Than Companies Realize?5
  • 6.
    How it alreadyhas gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 6
  • 7.
    How it alreadyhas gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 7
  • 8.
    Sugar bananas byMaksym Kozlenko licensed as CC-BY-SA-4.0
  • 9.
    The Cautious Pathto Strate- gic Advantage: How Militaries Should Plan for AI by Peter Eckersley
  • 10.
    Chinese facial recognitionsystem confuses bus ad with a jaywalker by Jon Fingas
  • 11.
    Turkish is agender neutral language by Alex Shams (28.11.2017) 11
  • 12.
    Google is fixinggender bias in its Translate service by Ivan Mehta 12
  • 13.
    Google’s technology willnot suggest gender-based pronouns because the risk is too high that its “Smart Compose” technology might predict someone’s sex or gender identity incorrectly and offend users, product leaders ... 4 4 Fearful of bias, Google blocks gender-based pronouns from new AI tool by Paresh Dave13
  • 14.
    The Machine FiredMe - No human could do a thing about it! Necessary orders are sent automatically and each order completion triggers another order. 4 idiallo.com/blog/when-a-machine-fired-me by Ibrahim Diallo14
  • 15.
    Responses to Critiqueson Machine Learning of Criminality Perceptions by Xiaolin Wu, Xi Zhang 15
  • 16.
    How it alreadyhas gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 16
  • 17.
    Racing Stripes CarTop View by qubodup original art: Autonomous Trap 001 by James Bridle (2017) 17
  • 18.
    The Moral Machineexperiment by Edmond Awad, Sohan Dsouza, Richard Kim, Jonathan Schulz, Joseph Henrich, Azim Shariff, Jean-François Bonnefon & Iyad Rahwan18
  • 19.
  • 20.
    Deep Neural Networksare Easily Fooled: High Confidence Predictions for Unrecognizable Images. by Nguyen A, Yosinski J, Clune J. at Evolving AI Lab, University of Wyoming 20
  • 21.
    Explaining and harnessingadversarial examples. by Goodfellow, Ian J., Jonathon Shlens, and Christian Szegedy. at OpenAI 21
  • 22.
    Strike (with) aPose: Neural Networks Are Easily Fooled by Strange Poses of Familiar Objects by Michael A. Alcorn, Qi Li, Zhitao Gong, Chengfei Wang, Long Mai, Wei-Shinn Ku, Anh Nguyen 22
  • 23.
    Fooling Neural Networksin the Physi- cal World with 3D Adversarial Objects by Anish Athalye, Logan Engstrom, Andrew Ilyas & Kevin Kwok at LabSix
  • 24.
    Adversarial Reprogramming ofNeural Networks. by Gamaleldin F. Elsayed, Ian Goodfellow and Jascha Sohl-Dickstein at Google Brain 24
  • 25.
    Robust Physical-World Attackson Deep Learning Models by Kevin Eykholt, Ivan Evtimov, Ear- lence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, Dawn Song
  • 26.
    Accessorize to aCrime: Real and Stealthy Attacks on State-of-the-Art Face Recognition by Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, Michael K. Reiter
  • 27.
  • 28.
    Malware: DeepLocker - ConcealingTargeted Attacks with AI Locksmithing DEF CON 25 - Hyrum Anderson - Evading next gen AV using AI Intrusion Detection: Here’s The Truth, GANs Easily Can Fool Intrusion Detection Systems Anti-Virus: G Data uses nerual network based DeepRay for detecting malware 5 5 Antiviren-Software mit KI-Technik: G Data setzt auf DeepRay28
  • 29.
    Fake fingerprints canimitate real ones in biometric systems – research by Philip Bontrager, Aditi Roy, Julian Togelius, Nasir Memon, Arun Ross
  • 30.
  • 31.
    America’s biggest body-cameracompany says facial recognition isn’t accurate enough for policing decisions 5 America’s biggest body-camera company says facial recognition isn’t accurate enough for policing decisions by Dave Gershgorn31
  • 32.
    Face Recognition FieldTest at Südkreuz 12 by C.Suthorn under CC-BY-SA-4.0
  • 33.
    While the authorsapplaud the force for attempting to develop an ethically sound and legally compliant approach to predictive policing, they warn that the ethical principles in the proposal are not developed enough to deal with the broad challenges this kind of technology could throw up, and that “frequently the details are insufficiently fleshed out and important issues are not fully recognized.” 6 6 Britain Is Developing an AI-Powered Predictive Policing System33
  • 34.
    How it alreadyhas gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 34
  • 35.
    Gender Shades by JoyBuolamwini (2018) and her MIT group
  • 36.
    Google Photos, y’allfucked up by Jacky Alcine (28.06.2015) 36
  • 37.
    Amazon’s Disturbing Planto Add Face Surveillance to Your Front Door37
  • 38.
    Amazon’s Face RecognitionFalsely Matched 28 Members of Congress With Mugshots 38
  • 39.
    Amazon’s Face RecognitionFalsely Matched 28 Members of Congress With Mugshots39
  • 40.
    I tested 14sentences for "perceived toxicity" using Perspectives. jessamyn west (25.08.2017) 40
  • 41.
    Design Justice, A.I.,and Escape from the Matrix of Domination by Sasha Costanza-Chock 41
  • 42.
    Amazon.com Inc’s machine-learning specialistsuncovered a big problem: their new recruiting engine did not like women. 6 Reuters - RPT-INSIGHT-Amazon scraps secret AI recruiting tool that showed bias against women42
  • 43.
  • 44.
    ... predictive modelsreinforce existing police practices ... ... patterns of police records, not patterns of crime ... ... cannot predict patterns of crime that are different ... 7 7 USA by Human Rights Data Analysis Group44
  • 45.
    ... the differencesin arrest rates by ethnic group between predictive policing and standard patrol practices were not statistically significant, ..." 8 8 Field-data Study Finds No Evidence of Racial Bias in Predictive Policing (2018) by Forensic Magazine45
  • 46.
  • 47.
    Positive Positive: equalchance False Positives/Negatives: unfair 9 9 How to Fight Bias with Predictive Policing (2018) by Eric Siegel in Scientific American47
  • 48.
    ... COMPAS isno more accurate or fair than predictions made by people with little or no criminal justice expertise. 10 10 The accuracy, fairness, and limits of predicting recidivism (2018) by Julia Dressel and Hany Farid in Science Advances48
  • 49.
    ... despite COMPAS’scollection of 137 features, the same accuracy can be achieved with a simple linear classifier with only two features.11 11 The accuracy, fairness, and limits of predicting recidivism (2018) by Julia Dressel and Hany Farid in Science Advances49
  • 50.
    The “bias” comesfrom base rates. 12 12 AI Ethics, Impossibility Theorems and Tradeoffs50
  • 51.
    Faceception ...recognizing “High IQ”,“White-Collar Offender”, “Pedophile”, and “Terrorist” ... According to Social and Life Science research personalities are affected by genes. Our face is a reflection of our DNA. 13 13 Faception51
  • 52.
    Predictive Staatsschutz GES-3D -Multi-Biometrische Gesichtserkennung MisPel-Projekt - Multi-Biometrisierte Forensische Personensuche in Lichtbild- und Videomassendaten INTEGER - Visuelle Entscheidungsunterstützung bei der Auswertung von Daten aus sozialen Netzwerken PANDORA - Propaganda, Mobilisierung und Radikalisierung zur Gewalt in der virtuellen und realen Welt X-SONAR - Extremistische Bestrebungen in Social Media Netzwerken RADIG-Z -Radikalisierung im digitalen Zeitalter RISKANT - Risikoanalyse bei islamistisch motivierten Tatgeneigten Survey: Predictive Staatsschutz auf Telepolis von Matthias Becker 52
  • 53.
    How it alreadyhas gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 53
  • 54.
    Everbody has adeep learning paper Thousands of papers ( 20.000/year) big part have low quality faked results or just lack of knowledge 54
  • 55.
    Agenda How it alreadyhas gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion55
  • 56.
    How it alreadyhas gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 56
  • 57.
    Quadrants Intended Unintended Inside BlackHat AI Bias a.k.a Vendor/User killer robots bad validation Cheating14 wrong use of model Outside Adversarial Attacks Bias a.k.a Attacker/User data poisoning 14 US border agents hacked their “risk assessment” system to recommend detention 100% of the time57
  • 58.
  • 59.
    Accountability - Ascenario build an AI for fraud detection for a bank AI blocks tranaction of company A A goes bankrupt A sues the bank developer (you) is questioned as a witness Can you provide answers? 15 15 Hype-Tech by Felix von Leitner (translated from german)59
  • 60.
    Accountability - andhow? reporting process? transparency to the public? power to decide on necessary changes? provability of work afterwards? 60
  • 61.
    Cost of MisbehavingAI Why should the vendor change something? legal consequences ethical/morally behaviour loss of reputation loss of opportunities loss of money 61
  • 62.
    Cost of MisbehavingAI Why should the users demand change? loss of money loss of security loss of freedom 62
  • 63.
    Defining and Researching"The Game" Defending is always harder than attacking 16 AI for/against: Detection Prediction Prevention Response 16 Is attacking machine learning easier than defending it?63
  • 64.
    Regulation - GDPR "Rightto be forgotten/"Right to erasure" Älgorithmic Fairnessänd "The Right to Explanation" 64
  • 65.
    Regulation - Understanding WhiteHouse report: Preparing for the future of Artificial Intelligence House of Lords report: AI in the UK: ready, willing and able? Bundestag then: some talk and a list of experts Bundestag now: KI-Eckpunktepapier 65
  • 66.
    How it alreadyhas gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 66
  • 67.
    Specification Design - bugs,ambiguities, side effects, high level language, preference learning, design protocols Emergent - wireheading, delusions, meta-Learning and sub-agents, detecting emerging behavior, reward hacking 16 Building safe artificial intelligence: specification, robustness, and assurance by DeepMind Safety Research - Pedro A. Ortega, Vishal Maini, and the DeepMind safety team67
  • 68.
    How it alreadyhas gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 68
  • 69.
    Robustness Risk and prevention- limitation, risk sensitivity, uncertainty estimates, safety margins, safe exploration, adversaries, cautious generalisation, verification Recovery and stability - availability, instability, error correction, fail-safe, distributional shift, graceful degradation 16 Building safe artificial intelligence: specification, robustness, and assurance by DeepMind Safety Research - Pedro A. Ortega, Vishal Maini, and the DeepMind safety team69
  • 70.
    Robustness - Adversaries image,video, malware, health records, ... Find, investigate, train on and robustify against attack vectors. defenses tend to be broken easily learn the adversaries as counter examples use regulization cryptography 17 17 Defense against adversarial attacks using machine learning and cryptography by Ingrid Fadelli70
  • 71.
    Robustness - Availability Availabilityof the processing? Can I DOS a Neural Network? Availability of predcitions or decisions? Is a poisoning attack a DOS? 71
  • 72.
    How it alreadyhas gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 72
  • 73.
    Ethics - discussionalapproach include other groups than engineers use checklists, like deon - An ethics checklist for data scientists there is no, I repeat NO, programmable ethics! ethics is a discussion and by this an ongoing process 73
  • 74.
    Ethics - technicalapproach enforce consent, clarity, consistency, control, consequences 18 18 Care about AI ethics? What you can do, starting today by Steven Adler74
  • 75.
    Chris Anderson: “withenough data, the numbers speak for themselves.” 18 The Hidden Biases in Big Data by Kate Crawford75
  • 76.
    Kate Crawford: ẞadly,they can’t. Data and data sets are not objective; they are creations of human design." 18 The Hidden Biases in Big Data by Kate Crawford76
  • 77.
    Fairness - Definition Anti-classification Classificationparity Calibration: If an algorithm produces a “score,” that “score” should mean the same thing for different groups. 77
  • 78.
    Fairness - Tools Aequitas- open source bias audit toolkit IBM - AI Fairness 360 Facebook creates internal tool - Fairness Flow "Microsoft is creating an oracle for catching biased AI algorithms" Tool: What-If tool deon - An ethics checklist for data scientists Fix bias in AI with more tech (probably AI) sigh 78
  • 79.
    How it alreadyhas gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 79
  • 80.
    Assurance Monitoring - interpretability,behavioural screening, traces, estimates of causal influence, tripwires & honeypots Enforcement - testability, interuptibility, boxing, physical security, encryption, signatures, authorisation, human override 18 Building safe artificial intelligence: specification, robustness, and assurance by DeepMind Safety Research - Pedro A. Ortega, Vishal Maini, and the DeepMind safety team80
  • 81.
    Assurance - Testing whynot currently? what to build? what to test? continuous regression tests? special items? biases? 81
  • 82.
    Assurance - PhysicalSecurity A neural network is some files on hardware. Can be copied, stolen, modified, ... 82
  • 83.
    Assurance - Interpretability TheMythos of Model Interpretability 19 expect superhuman standards, we only tend to have (post-rationalization) but needed for Debugging, Verification, Fairness, Transparency 19 The Mythos of Model Interpretability by Zachary C. Lipton83
  • 84.
    Assurance - Interpretability intrinsicvs. posthoc outcome: statistics, visualization, weights, data points, intrinsical model model-specific vs. model agnostic local vs. global remind the the needs and the knowledge of the different roles 84
  • 85.
    Introduction to LocalInterpretable Model-Agnostic Explanations (LIME) (2016) by Marco Tulio Ribeiro, Sameer Singh, Carlos Guestrin in O’Reilly
  • 86.
    Introduction to LocalInterpretable Model-Agnostic Explanations (LIME) (2016) by Marco Tulio Ribeiro, Sameer Singh, Carlos Guestrin in O’Reilly
  • 87.
    Question: I havean unexplainable AI that detects 1000 tumors and an explainable one that detects 100. Which one do I prefer? 87
  • 88.
    How it alreadyhas gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 88
  • 89.
    Development ETL - biases,fairness, privacy, data poisoning, ... Reproducibility - model management, data management, result management, ... 89
  • 90.
    In many real-worldcases, the researcher won’t have made notes or remember exactly what she did, so even she won’t be able to reproduce the model. 20 20 The Machine Learning Reproducibility Crisis (2018) by Pete Warden90
  • 91.
    Yet AI researcherssay the incentives are still not aligned with reproducibility. 21 21 Missing data hinder replication of artificial intelligence studies (2018) by Matthew Hutson in Science91
  • 92.
    Privacy - Confidentiality Privacy Encryption Possibletools: Differential Privacy Homomorphic Encryption Federated learning 22 22 Federated Learning: Collaborative Machine Learning without Centralized Training Data92
  • 93.
    How it alreadyhas gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 93
  • 94.
    I would say:It’s not the TÜV! 94
  • 95.
  • 96.
    How it alreadyhas gone wrong - some Examples Specification Robustness Ethics & Fairness Reproducibility What now? Frame Specification Robustness Ethics & Fairness Assurance Development Certificates/Verification Conclusion 96
  • 97.
    Education Educate AIbasics as part of data literacy in school and college. 97
  • 98.
    What can youdo? Educate Warn Support Research Develop 98
  • 99.
    Without our trust AIwill grow regardlessly. 99
  • 100.
    With the statedadvancements AI will have our trust and may work like expected. 100
  • 101.
    Thank you Alexander Pospiech DataEngineer inovex GmbH Lindberghstraß e 3 80939 München alexander.pospiech@inovex.de +49 153 - 31 81 051
  • 102.
    Conferences and MeetingsI Specific on the Dark Sides: Conference on Fairness, Accountability, and Transparency FATML - Fairness, Accountability, and Transparency in Machine Learning Interpretable ML Symposium @NIPS NIPS 2017 Tutorial - Fairness in Machine Learning Reproducibility in ML Workshop, ICML’18 IEEE 1st Deep Learning and Security Workshop Data Ethics workshop, KDD 2014 MAKE-Explainable AI Advances on Explainable Artificial Intelligence 102
  • 103.
    Conferences and MeetingsII Nemesis ’18 - 1st Workshop on Recent Advances in Adversarial Machine Learning NIPS 2018 Workshop on Security in Machine Learning Generic on AI: AI for Good Global Summit 103
  • 104.
    Conferences and MeetingsI General on Security: CCC DefCon SHA BlackHat 104
  • 105.
    Research Groups andOrganizations I AI specific: AINow - A research institute examining the social implications of artificial intelligence Evolving AI Lab, University of Wyoming OpenAI LabSix EFF on Artificial Intelligence & Machine Learning EFF - AI Progress Measurement EvalAI - Evaluating state of the art in AI EvadeML - Machine Learning in the Presence of Adversaries Adversarial Machine Learning, Università degli Studi di Cagliari 105
  • 106.
    Research Groups andOrganizations II SunBlaze at UCB Diskriminierung durch KI (Künstliche Intelligenz) (DiKI) Algorithmische Gegenmacht Center for Human-Compatible AI Technische und rechtliche Betrachtungen algorithmischer Entscheidungsverfahren by GI Algorithmenethik 106
  • 107.
    Research Groups andOrganizations I General: Trusted AI - IBM Research is building and enabling AI solutions people can trust Ethics/Fairness: SIENNA: Technology, ethics and human rights Human Rights Data Analysis Group AlgorithmWatch Netzpolitik on Predictive Policing 107
  • 108.
    Projects I Interpretability: Project: DARPAsExplainable Artificial Intelligence Project: Explain AI Project: DALEX: Descriptive mAchine Learning EXplanations 108
  • 109.
  • 110.
    Classes I CS 294:Fairness in Machine Learning, UC Berkeley 18739 Security and Fairness of Deep Learning, Carnegie Mellon Adversarial and Secure Machine Learning IEEE’s Artificial Intelligence and Ethics in Design The Ethics and Governance of Artificial Intelligence Attacking Networks with Adversarial Examples CS 294-149: Safety and Control for Artificial General Intelligence (Fall 2018) Machine-learning - crash-course -fairness 110
  • 111.
    Themensammlung I Netzpolitik onPredictive Policing EFF on Artificial Intelligence & Machine Learning EFF - AI Progress Measurement EvalAI - Evaluating state of the art in AI AI safety resources Ethics: AI Ethics Resources 111
  • 112.
    Github with ListsI General: Awful AI Ethics: Data Science and Ethics Resources Global Data Ethics Pledge (GDEP) ML and Security Machine Learning for Cyber Security Awesome Awesome AI Security Adversarial Attacks: Awesome Adversarial Examples for Deep Learning Awesome Adversarial Machine Learning 112
  • 113.
    Github with ListsII Awesome Machine Learning for Cyber Security Privacy; awesome-ai-privacy Private-Ai-Resources Machine Learning Ethics References Fairness: Toward ethical, transparent and fair AI/ML: a critical reading list for engineers, designers, and policy makers Interpretability; Awesome Interpretable Machine Learning Awesome awesome-machine-learning-interpretability Awesome 113
  • 114.
    Github with CodeI Ethics: deon - An ethics checklist for data scientists An Open Standard for Ethical Enterprise-Grade AI Interpretability: H20.ai: Machine Learning Interpretability (MLI) Explanation Explorer Interpretable Machine Learning with Python iml: interpretable machine learning ML Insights Fairness: Comparing fairness-aware machine learning techniques. 114
  • 115.
    Github with CodeII Themis ML - Fairness-aware Machine Learning Adversarial Attacks: Introduction to Adversarial Machine Learning DeepFool CleverHans EvadeML-Zoo AdvFlow Evaluating and Understanding the Robustness of Adversarial Logit Pairing IBM adversarial-robustness-toolbox deep-pwning FoolBox, 115
  • 116.
    Papers I LIME, LocalInterpretable Model-Agnostic Explanations (2016) 116
  • 117.
    Studies I European expertgroup seeks feedback on draft ethics guidelines for trustworthy artificial intelligence 117
  • 118.
    Blogs I a blogabout security and privacy in machine learning MLSec covert.io security + big data + machine learning Data Driven Security Automating OSINT BigSnarf Security of Machine Learning 118
  • 119.
    Slides I General: Hype-Tech byFelix von Leitner Adversarial Attacks: How CLEVER is your neural network? Robustness evaluation against adversarial examples by Pin-Yu Chen at IBM Research AI Privacy: Machine Learning & Privacy: It’s Complicated Fairness: AI Ethics, Impossibility Theorems and Tradeoffs 119
  • 120.
    Blog Posts I General: 6core falsehoods about the digital sphere Nein, Ethik kann man nicht programmieren You created a machine learning application. Now make sure it’s secure. Tooling: Deep automation in machine learning Fainess: Help Wanted: An Examination of Hiring Algorithms, Equity, and Bias (December 2018) Security: Attacks against machine learning — an overview 120
  • 121.
    Blog Posts II MachineLearning for Cybersecurity 101 Interpretability: Testing machine learning explanation techniques Explainable AI and the Legality of Autonomous Weapon Systems Adversarial Attacks: The Definitive Security Data Science and Machine Learning Guide Discussion and Survey of Adversarial Examples and Robustness in Deep Learning A Brief Introduction to Adversarial Examples Adversarial Sample Generation: Making Machine Learning Systems Robust for Security Survey on Security and Privacy of Machine Learning Attacking Machine Learning Detectors: the state of the art review121
  • 122.
    Podcasts I Fairness: O’Reilly DataShow Podcast - Why it’s hard to design fair machine learning models Privacy: How privacy-preserving techniques can lead to more robust machine learning models 122
  • 123.
    Videos - generalI General: Youtube: Stephen Fry describing our future with artificial intelligence and robots 34c3 - Beeinflussung durch Künstliche Intelligenz 34c3 - Deep Learning Blindspots SHA2017 - The Security and Privacy Implications of AI and Machine Learning Youtube - DEF CON 24 - Clarence Chio - Machine Duping 101: Pwning Deep Learning Systems Youtube: Do You Trust This Computer? TED - The era of blind faith in big data must end Managing Risk in Machine Learning - Ben Lorica (O’Reilly Media) 123
  • 124.
    Videos - generalII Ethics: Ethik der Algorithmen - Tom Hillenbrand zum "Schreckensszenario KI" Interpretability: [HUML16] 06: Zackary C. Lipton, The mythos of model interpretability "Why Should I Trust you?" Explaining the Predictions of Any Classifier, KDD 2016 Interpretable Machine Learning Using LIME Framework - Kasia Kulma (PhD), Data Scientist, Aviva TWiML - Trust in Prediction of Machine Learning Models - EMEA Meetup #4 - December 2018 Explaining Complex Machine Learning Models With LIME - Shirin Glander (codecentric AG) 124
  • 125.
    Videos - generalIII Fairness: Tutorial: 21 fairness definitions and their politics Adversarial Examples: Is "Adversarial Examples" an Adversarial Example? How to Successfully Harness AI to Combat Fraud and Abuse - RSA 2018 Adversarial Robustness, Theory and Practice Accountability: Lack of Accountability in Data Science: Why We Should All Be Building Inclusive AI 125
  • 126.
    Adversarial Examples -Examples I Audio Adversarial Examples 126
  • 127.
    Adversarial Attack CompetitionsI MNIST Adversarial Examples Challenge NIPS 2017 Competition: Non-targeted Adversarial Attack Introducing the Unrestricted Adversarial Examples Challenge 127
  • 128.
    Lists I Reward Hacking Specificationgaming examples in AI 128
  • 129.
    News I Autonomous Driving Uberself-driving car kills a pedestrian Incidents with Tesla Autopilot Even the cute Waymo 129
  • 130.
    Online Books I InterpretableMachine Learning Fairness and machine learning 130