Paper demonstrates digital-locker backdoors embedding encrypted neural subnets in host weights, unlocked by secret triggers, resisting SFT, weight noising, gradient trigger search, and latent adversarial training on chemical reaction proxy task.
Paper demonstrates digital-locker backdoors embedding encrypted neural subnets in host weights, unlocked by secret triggers, resisting SFT, weight noising, gradient trigger search, and latent adversarial training on chemical reaction proxy task.