When Does Machine Learning FAIL? Generalized Transferability for Evasion and Poisoning Attacks
summary
Defines the FAIL model (Features, Algorithm, Instances, Leverage) to quantify realistic attacker knowledge and control, and implements StingRay, a targeted poisoning procedure that crafts inconspicuous poisons without controlling labels. Evaluated against 4 applications using 3 learning algorithms, bypassing 2 defenses; white‑box success ≈50% while limited knowledge cuts success (existing models overestimate by ~36%). Empirical note: 89% of crafted Drebin candidates are labeled benign on VirusTotal, and StingRay proves practical.
AI-written summary · about
chapters (5)
related talks
28mSecurity of Machine Learning
42mDefense Against the Dark Arts and Potions: AI Models Can Be Easily Poisoned
46mAI in a Minefield: Learning from Poisoned Data
45mWinning by deception: tactics, techniques, and procedures of adversarial A.I.
40mEvasion, Poisoning, Extraction, and Inference: Tools to Defend and Evaluate
52mSleeping With One AI Open: An Introduction to Attacks Against...
45mSecuring AI: There Is No Try, Only Do!
20mMachine Learning Attacks – A New Era of Security Epidemic