🤖 AI Summary
To address the severe utility degradation of large language models (LLMs) arising from the combined application of differential privacy (DP) and model compression, this paper proposes DistilDP—a novel framework that jointly optimizes privacy preservation and model efficiency. Under a stringent DP budget of ε = 2, DistilDP leverages a differentially private teacher model to generate high-quality synthetic text and introduces a three-way knowledge distillation strategy: hard-label classification, soft-label probability distribution matching, and hidden-layer representation alignment—enabling multi-granularity knowledge transfer. To our knowledge, this is the first work to deeply integrate DP-compliant synthetic data generation with multi-granularity distillation. Evaluated on the Big Patent dataset, the distilled student model achieves ≥9.0% reduction in perplexity, substantially outperforming existing privacy-aware compression baselines. DistilDP establishes a verifiable, scalable paradigm for privacy-sensitive LLM lightweighting.
📝 Abstract
Large Language models (LLMs) are achieving state-of-the-art performance in many different downstream tasks. However, the increasing urgency of data privacy puts pressure on practitioners to train LLMs with Differential Privacy (DP) on private data. Concurrently, the exponential growth in parameter size of LLMs necessitates model compression before deployment of LLMs on resource-constrained devices or latency-sensitive applications. Differential privacy and model compression generally must trade off utility loss to achieve their objectives. Moreover, simultaneously applying both schemes can compound the utility degradation. To this end, we propose DistilDP: a novel differentially private knowledge distillation algorithm that exploits synthetic data generated by a differentially private teacher LLM. The knowledge of a teacher LLM is transferred onto the student in two ways: one way from the synthetic data itself -- the hard labels, and the other way by the output distribution of the teacher evaluated on the synthetic data -- the soft labels. Furthermore, if the teacher and student share a similar architectural structure, we can further distill knowledge by aligning the hidden representations between both. Our experimental results demonstrate that DistilDP can substantially improve the utility over existing baselines, at least $9.0$ PPL on the Big Patent dataset, with strong privacy parameters, $epsilon=2$. These promising results progress privacy-preserving compression of autoregressive LLMs. Our code can be accessed here: https://github.com/james-flemings/dp_compress.