Privacy Preserving and Efficient Machine Learning Algorithms

Efstathia Soufleri

Keywords: Differential Privacy, Synthetic Data, Image Classification, Model Compression, Mixed-Precision Quantization, Deep Neural Networks (DNNs), Compressed Video, Action Recognition, Knowledge Distillation, Privacy-Preserving Machine Learning, Efficient Deep Learning, Neural Network Optimization, Computer Vision

TL;DR: A privacy-preserving and efficiency-focused study introducing methods for synthetic data generation, neural network compression, and action recognition on compressed videos.

Abstract:

This dissertation addresses two key challenges in modern deep learning: data privacy and computational efficiency. As deep learning expands across domains like image classification and natural language processing, privacy restrictions often limit access to sensitive datasets, while the growing complexity of deep models strains deployment resources.

To tackle the privacy issue, we introduce DP-ImgSyn, a novel framework for differentially private image synthesis. It generates synthetic images aligned with a private dataset by optimizing public image initializations using batch norm statistics from a DP-trained teacher network. The resulting images (1) guarantee differential privacy, (2) preserve utility for training classifiers, and (3) are visually dissimilar from private data.

On the efficiency front, we explore layer-wise mixed-precision quantization to compress DNNs. We train a Multi-Layer Perceptron to predict optimal bit-widths per layer using the KL divergence between full-precision and quantized model outputs, reducing the search complexity while maintaining accuracy.

Finally, we propose PKD (Progressive Knowledge Distillation) and WISE (Weighted Inference with Scaled Ensemble) for efficient action recognition on compressed videos. Our method processes motion vectors, residuals, and intra-frames directly, avoiding costly decompression. Knowledge is progressively distilled across modalities using internal classifiers, and predictions are fused via a learned weighted ensemble, improving both speed and performance.

Together, these contributions advance the development of scalable, privacy-preserving, and resource-efficient deep learning systems for real-world vision applications.

Submission Number: 109