Cihan Özhan Founder of DeepLab, http://deeplab.co

DeepLab : Technologies ML/DL/CV Based • Technologies – Go, Python, C/C++, Rust, C# – PyTorch, TensorFlow, Keras, scikit-learn • As Web, Mobile, IoT/Edge ve Back-End… – OpenCV – … and dozens of different tools and equipment… • Cloud Computing – AWS Machine Learning – Google Cloud Machine Learning – IBM Watson Machine Learning – Microsoft Azure Machine Learning – … and various cloud solutions… • Distributed Systems – Distributed Databases – Distributed Deep Learning

AI Security Machine Learning, Deep Learning and Computer Vision Security Cihan Özhan | Founder of DeepLab | Developer, AI Engineer, AI Hacker, Data Master

OWASP Top 10 List OWASP (Open Web Application Security Project)

No content

AI Data Objects • Image • Text • File • Voice • Video • Data • 3D Object

ML/DL Applications • Image Classification • Pose Estimation • Face Recognition • Face Detection • Object Detection • Question Answering System • Semantic Segmentation • Text Classification • Text Recognition • Sentiment Analysis • Industrial AI • Autonomous Systems • and more…

ML/DL Algorithms • Classification (Supervised) • Clustering (Unsupervised) • Regression (Supervised) • Generative Models (Semi-Supervised) • Dimensionality Reduction (Unsupervised) • Reinforcement Learning (Reinforcement)

MLaaS? Machine Learning as a Service MLaaS is the method in which ML/DL algorithms and software are offered as a component of cloud computing services. MLaaS = (SaaS + [ML/DL/CV])

Hidden Technical Debt in Machine Learning Systems https://papers.nips.cc/paper/2015/file/86df7dcfd896fcaf2674f757a2463eba-Paper.pdf Usually the whole team, focus and area where resources are directed!

Model Lifecycle Machine Learning Model Development Lifecycle

Model Lifecycle Machine Learning Model Development Lifecycle We start here! ML model preparation process The chore but the imperative: Preparing the data! We prepared the model! We train the model with data. Cloud or On-Premise We tested the trained model with test data! The trained model is packaged for the programmatic environment. Post release: The model is constantly monitored.

How to Publish a Machine Learning project?

MartinFowler.com

MartinFowler.com

Machine Learning Security

Basic Security Issues Intentional Issues Unintentional Issues Perturbation Attack Reward Hacking Poisoning Attack Side Effects Model Inversion Distributional Shifts Membership Inference Natural Adversarial Examples Model Stealing Common Corruption Reprogramming ML system Incomplete Testing Adversarial Example in Pyhsical Domain Malicious ML provider recovering training data Attacking the ML supply chain Backdoor ML Exploit Software Dependencies

No content

Adversarial Attack : Image (https://adversarial.io/)

No content

https://openai.com/blog/adversarial-example-research/

https://hackernoon.com/adversarial-attacks-how-to-trick-computer-vision-7484c4e85dc0

Adversarial Attack : Speech-to-Text (https://people.eecs.berkeley.edu/~daw/papers/audio-dls18.pdf)

https://arxiv.org/pdf/2006.03575.pdf

Adversarial Attack : NLP https://arxiv.org/pdf/2005.05909.pdf https://github.com/QData/TextAttack

No content

Adversarial Attack : Remote Sensing (https://arxiv.org/pdf/1805.10997.pdf)

Adversarial Attack : Satellite (https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8823003)

Adversarial Attack : Military https://spectrum.ieee.org/automaton/artificial-intelligence/embedded-ai/adversarial-attacks-and-ai-systems

Adversarial Attack : Military https://www.sto.nato.int/publications/STO%20Meeting%20Proceedings/STO-MP-IST-160/MP-IST-160-S1-5.pdf

Adversarial Attack : Autonomous Driving https://web.cs.ucla.edu/~miryung/Publications/percom2020-autonomousdriving.pdf https://github.com/ITSEG-MQ/Adv-attack-and-defense-on-driving-model

Security Research of Tesla Autopilot (40 Pages) https://keenlab.tencent.com/en/whitepapers/Experimental_Security_Research_of_Tesla_Autopilot.pdf

Industrial AI and Autonomous Systems - Cihan Özhan (Turkish) https://www.youtube.com/watch?v=ID_tw5iq6Xs

https://medium.com/@ml.at.berkeley/tricking-neural-networks-create-your-own-adversarial-examples-a61eb7620fd8

Exploit Software Dependencies • It takes advantage of the vulnerabilities of the software the system is connected to, not algorithms. • Prevention: – Security Scan – Security Reports – Be Careful : Wrappers ve Pre-Build Environment – Use Less Dependency – Dependency Management Tools • Synk : Synk.io • Python Poetry : python-poetry.org • Bandit : – Bandit is a tool designed to find common security issues in Python code. – https://github.com/PyCQA/bandit • pyup.io/safety • requires.io – etc…

Tool/Library Security (TensorFlow) • TensorFlow (tools like) is designed for internal communication, not for running on untrusted networks. • These tools (ModelServer etc.) do not have built-in authorization. • It can read and write files, send and receive data over the network… • (!) TensorFlow Models as Programs • (!) Running Untrusted Models • (!) Accepting Untrusted Inputs https://github.com/tensorflow/tensorflow/blob/master/SECURITY.md

https://www.tensorflow.org/tutorials/generative/adversarial_fgsm

Cihan Özhan Links • cihanozhan.com • linkedin.com/in/cihanozhan • medium.com/@cihanozhan • youtube.com/user/OracleAdam • twitter.com/UnmannedCode • github.com/cihanozhan Contact • cihan@deeplab.co