Making Sense of the Landscape of Attacks and Defenses Against AI
summary
NCCoE and MITRE present a concrete, modular testbed for measuring attacks and defenses on image classifiers, claiming it makes combinatorial experiments repeatable and low-friction. It implements poisoning, evasion, oracle attack classes and defenses like data sanitization, feature squeezing and robust training. The architecture uses Docker containers + REST API + Redis + MLFlow with TensorFlow/PyTorch worker nodes to run experiments, store results and track ML runs. Code lands on GitHub; the demo runs and is intended for plug‑and‑play swapping of attacks, defenses and datasets.
AI-written summary · about
chapters (22)
- 0:00Intro
- 0:58Maior Advances in A.I. Continue to Drive Need for Universal Understanding of Risks
- 2:17Fundamental and Applied Research and Standards for Al Technologies (FARSAIT)
- 3:08Trustworthy Al's Foundation: From Technical Requirements to Policy Creation
- 6:29Core Building Blocks of Trustworthy Al
- 7:14A Taxonomy and Terminology of Adversarial Attacks and Defenses on Machine Learning (Draft NISTIR 8269)
- 8:08Overview of Attacks
- 9:32Visual Sampling af Attacks
- 11:39Overview of Defenses
- 12:38Visual Sampling of Defenses
- 15:02Challenge • Tame the wild landscape of attacks and defenses for ML models and algorithms
- 16:30Goals and Opportunities
- 18:49Choose Your Own Adventure
- 19:36Numerous Experiments Built From Common Plugins
- 21:28Testbed Architecture
- 23:07The Fundamental Units of a Testbed Experiment
- 24:02The Four User Levels of the Testbed
- 25:26Code Snippets for Different levels
- 26:19Evaluation Metrics
- 29:06Dataset Characteristics
- 31:03Many Deployment Scenarios
- 34:17Apply What You Have Learned Today
related talks
53mSecurity of Machine Learning Systems
25mThreat Modeling for AI Apps with Attacks as Code
18mCommit Secure AI: Embedding Security in Code, Pipelines, and Production
49mMeasuring the Difference: Metric Development at NCCoE’s Securing AI Testbed
39mTrust Cards For AI
52mSleeping With One AI Open: An Introduction to Attacks Against...
46mA Short Introduction to Hacking Artificial Intelligence
48mStay Ahead of Adversarial AI in OT/ICS Environments – Mitigating CWE-1039