This paper introduces DistilDP, a differentially private knowledge distillation algorithm that leverages synthetic text generation to improve the utility of compressed language models while maintaining privacy. The method trains a student model using knowledge distilled from a differentially private teacher model, which generates synthetic data. The teacher's knowledge is transferred to the student in two ways: through the synthetic data itself (hard labels) and through the teacher's output distribution (soft labels). Additionally, if the teacher and student have similar architectures, hidden representations can be aligned to further enhance performance. The proposed framework avoids applying differential privacy (DP) during the knowledge distillation process by using DP-synthetic data generated by the teacher. This approach reduces the computational and memory costs associated with DP-SGD, which is typically applied twice in existing methods (once for the teacher and once for the student). DistilDP achieves significant improvements in model utility, with results showing a 9.0 PPL reduction on the Big Patent dataset using strong privacy parameters (ε = 2). The method is evaluated on three datasets: Yelp, Big Patent, and DBpedia. It outperforms existing baselines, including a student model fine-tuned with DP-SGD, a student model fine-tuned using DPKD, and a student model trained only on DP synthetic data. The results demonstrate that aligning the output distributions of the teacher and student, as well as aligning hidden representations, are crucial for improving the student's performance. The paper also includes ablation studies showing that the distillation loss weighting parameter (λ) and temperature parameter (t) significantly affect performance. Increasing the number of synthetic text data also improves the student's performance. Additionally, the inclusion of an MSE loss on the hidden representations further enhances the student's utility. The framework is model-agnostic and does not require specific architectural assumptions between the teacher and student. However, aligning hidden representations can further improve performance. The method is efficient, requiring only one application of DP-SGD for the teacher, which reduces computational and memory costs compared to existing methods. The paper concludes that differentially private knowledge distillation for autoregressive large language models can be effective. The results show that synthetic data generation, combined with knowledge distillation, can significantly improve the utility of compressed models while maintaining privacy. The method is promising for privacy-preserving compression of large language models.This paper introduces DistilDP, a differentially private knowledge distillation algorithm that leverages synthetic text generation to improve the utility of compressed language models while maintaining privacy. The method trains a student model using knowledge distilled from a differentially private teacher model, which generates synthetic data. The teacher's knowledge is transferred to the student in two ways: through the synthetic data itself (hard labels) and through the teacher's output distribution (soft labels). Additionally, if the teacher and student have similar architectures, hidden representations can be aligned to further enhance performance. The proposed framework avoids applying differential privacy (DP) during the knowledge distillation process by using DP-synthetic data generated by the teacher. This approach reduces the computational and memory costs associated with DP-SGD, which is typically applied twice in existing methods (once for the teacher and once for the student). DistilDP achieves significant improvements in model utility, with results showing a 9.0 PPL reduction on the Big Patent dataset using strong privacy parameters (ε = 2). The method is evaluated on three datasets: Yelp, Big Patent, and DBpedia. It outperforms existing baselines, including a student model fine-tuned with DP-SGD, a student model fine-tuned using DPKD, and a student model trained only on DP synthetic data. The results demonstrate that aligning the output distributions of the teacher and student, as well as aligning hidden representations, are crucial for improving the student's performance. The paper also includes ablation studies showing that the distillation loss weighting parameter (λ) and temperature parameter (t) significantly affect performance. Increasing the number of synthetic text data also improves the student's performance. Additionally, the inclusion of an MSE loss on the hidden representations further enhances the student's utility. The framework is model-agnostic and does not require specific architectural assumptions between the teacher and student. However, aligning hidden representations can further improve performance. The method is efficient, requiring only one application of DP-SGD for the teacher, which reduces computational and memory costs compared to existing methods. The paper concludes that differentially private knowledge distillation for autoregressive large language models can be effective. The results show that synthetic data generation, combined with knowledge distillation, can significantly improve the utility of compressed models while maintaining privacy. The method is promising for privacy-preserving compression of large language models.