Adam Dziedzic

Membership inference attacks can exploit tabular ICL, but TabPATE offers a robust defense that preserves model utility without requiring public data.

Efficiently removing undesirable concepts from image generations without sacrificing quality could redefine how we manage biases in generative models.

Natural identifiers can transform LLM privacy audits by eliminating the need for retraining and inaccessible datasets, making post-hoc assessments practical and scalable.

Existing methods misclassify generated samples as training members, but the Data Circuit Breaker reliably distinguishes between the two, even in challenging contexts.

Distribution shifts can drastically elevate privacy risks in LLM adaptations, challenging the reliability of theoretical DP guarantees.