The Design and Development of a Game to Study Backdoor Poisoning Attacks: The Backdoor Game
summary
A browser game implements activation clustering to let players probe poisoned image classifiers, upload candidate backdoor images, and test triggers. Developers run eight semi‑structured interviews and a Mechanical Turk deployment with N=68 where 66% win, 191 guesses produce 39 correct photos and an average of seven peeks. The team moves a 3D cluster view to 2D for usability, documents cheating versus original submissions, and shows the platform both collects diverse backdoor objects and measures trigger effectiveness.
AI-written summary · about
related talks
59mBackdoors & Breaches: Live Tabletop Exercise Demo
01h 08mBackdoors & Breaches: Live Tabletop Exercise Demo
54mWant to Play a Game? Get Your Game On with Hacking Gamification
30mShall We Play a Game?
45mWinning by deception: tactics, techniques, and procedures of adversarial A.I.
43mGame of Hacks: The Mother of All Honeypots
40mBackdoor Investigation and Incident Response: From Zero to Profit
53mDoors of (AI)pportunity: The Front and Backdoors of LLMs