🤖 AI Summary
Generative models are prone to overfitting and memorizing rare training samples, leading to privacy leakage and inflated benchmark performance. Method: We propose a data-centric memorization analysis framework that—novelty—integrates sample difficulty (early-training loss) and forgetting event frequency to define a memorization score, enabling a four-quadrant classification scheme to identify memorization hotspots; we theoretically prove that the memorization score’s lower bound and uniform stability jointly bound the generalization gap, and accordingly design dynamic data reweighting and pruning strategies. Contribution/Results: Removing only 10% of training data reduces synthetic watermark extraction success rate by over 40% while increasing perplexity by less than 0.5, substantially mitigating information leakage with negligible degradation in generation quality. Our core innovation lies in an interpretable, theoretically grounded, data-level intervention paradigm.
📝 Abstract
Modern generative models risk overfitting and unintentionally memorizing rare training examples, which can be extracted by adversaries or inflate benchmark performance. We propose Generative Data Cartography (GenDataCarto), a data-centric framework that assigns each pretraining sample a difficulty score (early-epoch loss) and a memorization score (frequency of ``forget events''), then partitions examples into four quadrants to guide targeted pruning and up-/down-weighting. We prove that our memorization score lower-bounds classical influence under smoothness assumptions and that down-weighting high-memorization hotspots provably decreases the generalization gap via uniform stability bounds. Empirically, GenDataCarto reduces synthetic canary extraction success by over 40% at just 10% data pruning, while increasing validation perplexity by less than 0.5%. These results demonstrate that principled data interventions can dramatically mitigate leakage with minimal cost to generative performance.