Backdoor Detection
Overview
Backdoor detection studies how to find models or inputs that behave normally in ordinary use but produce an attacker-chosen response when a hidden trigger is present. Methods inspect poisoned training data, model weights and activations, or behavior at inference time; common approaches include spectral analysis, activation clustering, trigger inversion, controlled input perturbations, and runtime monitoring. Work spans image and text classifiers, language and multimodal models, audio, graphs, and reinforcement-learning agents, with evaluations on benchmarks that vary attack types, trigger visibility, and access to training data or model internals. A central challenge is making detectors transfer to unseen attacks and model families while keeping false alarms low, since adaptive backdoors can evade signals that work in closed-world tests.
Latest papers
53 papers8 papers in the last four weeks, up 100% on the four weeks before. 0.0% of all new papers.