mlsanityData to 14 Oct 2026

Characterizing the Training Dynamics of Private Fine-tuning with Langevin diffusion

arXiv 2402.18905 ↗First posted 29 Feb 2024cs.LG cs.AI cs.CRcited by 4

Not on the map. Why?

Abstract (arXiv)

We show that differentially private full fine-tuning (DP-FFT) can distort pre-trained backbone features based on both theoretical and empirical results. We identify the cause of the distortion as the misalignment between the pre-trained backbone and the randomly initialized linear head. We prove that a sequential fine-tuning strategy can mitigate the feature distortion: first-linear-probing-then-fine-tuning (DP-LP-FFT). A new approximation scheme allows us to derive approximate upper and lower bounds on the training loss of DP-LP and DP-FFT, in a simple but canonical setting of 2-layer neural networks with ReLU activation. Experiments on real-world datasets and architectures are consistent with our theoretical insights. We also derive new upper bounds for 2-layer linear networks without the approximation. Moreover, our theory suggests a trade-off of privacy budget allocation in multi-phase fine-tuning methods like DP-LP-FFT.

Key facts. Cited by 4 papers in the map (data through 14 Oct 2026). Builds on 37 of its 61 references that are in the map. First posted on arXiv 29 Feb 2024.

Builds on

37 of 61 references.

References · 37

  1. 2014ImageNet Large Scale Visual Recognition Challengecited by 8,498
  2. 2015Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shiftcited by 7,005
  3. 2020Bootstrap your own latent: A new approach to self-supervised Learningcited by 4,750
  4. 2020Training data-efficient image transformers & distillation through attentioncited by 4,586
  5. 2016Wide Residual Networkscited by 3,070
  6. 2018UMAP: Uniform Manifold Approximation and Projection for Dimension Reductioncited by 2,898
  7. 2016Deep Learning with Differential Privacycited by 2,728
  8. 2020Improved Baselines with Momentum Contrastive Learningcited by 2,287
  9. 2021An Empirical Study of Training Self-Supervised Vision Transformerscited by 1,616
  10. 2018Group Normalizationcited by 1,524
  11. 2022Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distributioncited by 739
  12. 2019A Theoretical Analysis of Contrastive Unsupervised Representation Learningcited by 530
25 more
  1. 2021Differentially Private Fine-tuning of Language Modelscited by 400
  2. 2018On the Optimization of Deep Networks: Implicit Acceleration by Overparameterizationcited by 284
  3. 2022Unlocking High-Accuracy Differentially Private Image Classification through Scalecited by 224
  4. 2020On the Theory of Transfer Learning: The Importance of Task Diversitycited by 186
  5. 2018Algorithmic Regularization in Learning Deep Homogeneous Models: Layers are Automatically Balancedcited by 177
  6. 2020Better Fine-Tuning by Reducing Representational Collapsecited by 157
  7. 2020Understanding Gradient Clipping in Private SGD: A Geometric Perspectivecited by 142
  8. 2021On the Role of Neural Collapse in Transfer Learningcited by 100
  9. 2022Neural Collapse: A Review on Modelling Principles and Generalizationcited by 95
  10. 2021Learning to See by Looking at Noisecited by 92
  11. 2022MiniViT: Compressing Vision Transformers with Weight Multiplexingcited by 88
  12. 2021Differential Privacy Dynamics of Langevin Diffusion and Noisy Gradient Descentcited by 73
  13. 2021What can linearized neural networks actually say about generalization?cited by 49
  14. 2021Global Convergence of Gradient Descent for Asymmetric Low-Rank Matrix Factorizationcited by 45
  15. 2023Abide by the Law and Follow the Flow: Conservation Laws for Gradient Flowscited by 45
  16. 2023Why Is Public Pretraining Necessary for Private Model Training?cited by 37
  17. 2021On the Convergence and Calibration of Deep Learning with Differential Privacycited by 36
  18. 2023Early Neuron Alignment in Two-layer ReLU Networks with Small Initializationcited by 28
  19. 2023DPZero: Private Fine-Tuning of Language Models without Backpropagationcited by 26
  20. 2021Differentially Private SGD with Non-Smooth Lossescited by 22
  21. 2023A Closer Look at Model Adaptation using Feature Distortion and Simplicity Biascited by 20
  22. 2023Differentially Private Image Classification by Learning Priors from Random Processescited by 19
  23. 2023Differentially Private SGD Without Clipping Bias: An Error-Feedback Approachcited by 14
  24. 2023DP-SGD Without Clipping: The Lipschitz Neural Network Waycited by 11
  25. 2023Initialization Matters: Privacy-Utility Analysis of Overparameterized Neural Networkscited by 10

Built on by

Cited by year

Citing papers per year: 2024: 2; 2025: 2.

4 papers.

  1. 2024DP-MemArc: Differential Privacy Transfer Learning for Memory Efficient Language Modelscited by 2
  2. 2025On Optimal Hyperparameters for Differentially Private Deep Transfer Learningcited by 1
  3. 2024Mitigating Noise Detriment in Differentially Private Federated Learning with Model Pre-trainingcited by 0
  4. 2025Lower Bounds for Public-Private Learning under Distribution Shiftcited by 0