Targeted suppression of non-literal copyright leakage

Develop targeted non-literal copyright-suppression methods that achieve fine-grained semantic discrimination rather than relying on broad pre-generation blocking, by using finer-grained representation-level classifiers that distinguish narrative engagement from verbatim copying or training signals that explicitly penalize semantic similarity.

Background

CopyShield evaluates contrastive decoding, Direct Preference Optimization, and activation intervention across literal leakage, non-literal semantic leakage, utility, and degeneracy. Although all three defenses reduce some forms of leakage, none provides targeted suppression of semantically reproducing responses. Activation intervention achieves the lowest non-literal flagging rate primarily by blocking most queries before generation, rather than by distinguishing harmful reproduction from legitimate narrative discussion; DPO focuses on verbatim continuation and leaves non-literal flagging near the undefended baseline.

The unresolved problem is therefore to design copyright defenses that suppress semantic reproduction selectively while preserving useful responses to legitimate questions and discussions. The paper identifies two concrete directions: representation-level classifiers with finer discrimination between narrative engagement and verbatim copying, and training objectives that explicitly penalize semantic similarity.