Research map / Large language models / Language, safety and interpretability

How to Protect Copyright Data in Optimization of Large Language Models?

AAAI 2024 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks · 18 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes "copyright regression," a modified softmax regression objective with an inverse term on copyrighted data, proves convexity and Hessian Lipschitzness, and runs synthetic Gaussian-data experiments to show a controllable gap in fit between copyrighted and other data.

Problem

Preventing LLMs trained on copyrighted data from reproducing that data, framed as a softmax regression problem underlying transformer training and attention.

Approach

Adds an inverse penalty term gamma_c * l1(x)^{-1} on the copyrighted-data loss to the softmax regression objective, making the copyrighted-data fit worse than the other data by a margin tau. The authors argue this is a plug-and-play training modification that needs no extra model and does not slow decoding, and they prove convexity and Hessian Lipschitzness of the objective.

Key results

• Experiments on synthetic Gaussian data (n=10000, d=512) show the copyright-vs-other error gap tau grows with gamma_c, consistent with the theoretical bound tau_c = sqrt(2 gamma_c)/n1 - eps2/n2.
• The authors report that even at n1=8000, MAE on copyrighted data remains above the Random baseline's error, which they interpret as protection against outputting copyrighted data.
• Convexity is claimed via Hessian positive definiteness under a weight condition on w_i^2, and Hessian Lipschitzness is shown with constant polynomial in n and exp(40R^2).

Measured effects reported in the paper

Surprising findings

• Increasing the proportion of copyrighted data n1/n shrinks the performance gap between copyrighted and other data, which weakens the protection.

Limitations

• Experiments only use synthetic Gaussian data and a linear softmax regression, not real LLM training or copyrighted text.
• The copyright-protection notion is a loss gap between copyrighted and other data, not a demonstrated reduction in verbatim regurgitation.
• The convexity and Lipschitz claims rely on bounded-norm assumptions (||x||, ||A|| bounded by R) and the paper's own definitions contain inconsistencies (e.g., the MSE and MAE definitions are swapped).

Experimental setup

Datasets: synthetic Gaussian data (n=10000, d=512)

Benchmarks: synthetic Gaussian regression setup

Metrics: MSE gap tau_MSE MAE gap tau_MAE MAE on copyrighted data

Baselines: Random

Builds on: softmax regression (attention as regression) near access-freeness exponential regression for attention

Closest papers by meaning

• Copyright-Protected Language Generation via Adaptive Model Fusion (ICLR 2025 · similarity 0.86)
• Can Watermarking Large Language Models Prevent Copyrighted Text Generation and Hide Training Data? (AAAI 2025 · similarity 0.86)
• Copyright Traps for Large Language Models (ICML 2024 · similarity 0.85)
• DP-OPT: Make Large Language Model Your Privacy-Preserving Prompt Engineer (ICLR 2024 · similarity 0.85)
• A Theory of Training Profit-Optimal LLMs (NeurIPS 2026 · similarity 0.84)
• Scaling Laws for Differentially Private Language Models (ICML 2025 · similarity 0.84)