Cascading Adversarial Bias from Injection to Distillation in Language Models

From Google TechTalks · 39:15

Show by Google TechTalks

A Google TechTalk, 2024-10-02, presented by Harsh ChauhariML Privacy Seminar. ABSTRACT: Our work examines how adversarial bias injection during training is not only propagated but significantly amplified when distilling language models, revealing a critical security vulnerability. With minimal data poisoning (just 0.25% of training data), researchers demonstrate that biases become more pronounced in student models than in their teachers—in targeted scenarios, student models generate biased content 76.9% of the time versus 69.4% in teachers, while untargeted biases appear up to 29.2x more frequently in student models on unseen tasks. Testing across various bias types, distillation methods, and data modalities, our study shows current defense mechanisms are inadequate against these sophisticated attacks, highlighting a need for specialized safeguards and proposing practical design principles for future mitigation strategies.Uploaded 26-01-26

More episodes of Google TechTalks

More episodes

PodSearch is a free, independent podcast search engine. Read about PodSearch, the privacy policy or contact us.