"Copyright-Protected Language Generation via Adaptive Model Fusion"

Smart model combination during inference time blocks LLMs from memorizing copyrighted content.

CP-Fuse, proposed in this paper, adaptively combines models trained on different copyright sets during inference to prevent LLMs from reproducing protected content while maintaining generation quality.

-----

https://arxiv.org/abs/2412.06619v1

🤔 Original Problem:

LLMs often reproduce copyrighted training data verbatim, leading to legal risks and potential lawsuits. Existing protection methods like differential privacy training are computationally expensive and degrade model performance.

-----

🔧 Solution in this Paper:

→ CP-Fuse combines outputs from multiple LLMs trained on disjoint sets of copyrighted material during inference time

→ The method adaptively aggregates model logits to minimize copyright content reproduction through a balancing property

→ When one model dominates generation, CP-Fuse automatically shifts weight to other models to prevent memorized content reproduction

→ The approach works post-hoc, allowing seamless integration with other protection techniques

-----

💡 Key Insights:

→ Model fusion can effectively prevent copyright violations without compromising generation quality

→ Adaptive weighting based on generation history is crucial for preventing memorization

→ Post-processing approaches can be more practical than training-time methods

→ The balancing property mathematically guarantees reduced regurgitation

-----

📊 Results:

→ Reduces copyrighted content reproduction by over 25x compared to baselines

→ Maintains equivalent code generation pass@1 scores as original models

→ Achieves same text generation fluency scores as unprotected models

→ Shows robustness against prompt-based extraction attacks