Backdoor Attacks

Backdoor attacks study how hidden behaviors can be implanted in machine-learning systems and activated by a trigger, even when performance on ordinary inputs appears normal. Attackers use poisoned training data, malicious model updates, or compromised components; triggers range from image patches and text styles to graph structures, timing patterns, and an agent’s internal state. Research spans image and speech classifiers, language and diffusion models, federated learning, reinforcement learning, and agent systems, with defenses including data filtering, model inspection, fine-tuning, and runtime monitoring. A central challenge is that defenses can miss unfamiliar or deployment-specific triggers, while stronger mitigation may reduce useful performance and some backdoors may be difficult to detect even with access to model parameters.