Token by Token, Compromised: Backdoor Vulnerabilities in Unified Autoregressive Models

Name: Token by Token, Compromised: Backdoor Vulnerabilities in Unified Autoregressive Models
Uploaded: 2026-10-10T17:05:53.539293+00:00
Duration: 777 s
Description: This work introduces Token by Token Backdoor Attack (ToBAC), a method that exploits unified autoregressive models to create multimodal backdoors. The attack uses 'autoregressive self-poisoning' where a trigger first generates a poisoned image, which then influences the model to produce malicious text. This allows triggers to work through both visual and textual inputs.

This work introduces Token by Token Backdoor Attack (ToBAC), a method that exploits unified autoregressive models to create multimodal backdoors. The attack uses 'autoregressive self-poisoning' where a trigger first generates a poisoned image, which then influences the model to produce malicious text. This allows triggers to work through both visual and textual inputs.

Open the full episode