DCT-CryptoNets: Scaling Private Inference in the Frequency Domain
Authors: Arjun Roy, Kaushik Roy
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | DCT-Crypto Nets demonstrates a substantial latency reductions of up to 5.3 compared to prior work on benchmark image classification tasks. Notably, it demonstrates inference on the Image Net dataset within 2.5 hours (down from 12.5 hours on equivalent 96-thread compute resources). Furthermore, by learning perceptually salient low-frequency information DCT-Crypto Nets improves the reliability of encrypted predictions compared to RGB-based networks by reducing error accumulating homomorphic bootstrap operations. DCT-Crypto Nets also demonstrates superior scalability to RGB-based networks by further reducing computational cost as image size increases. This study demonstrates a promising avenue for achieving efficient and practical private inference of deep learning models on high resolution images seen in real-world applications. |
| Researcher Affiliation | Academia | Arjun Roy Purdue University EMAIL Kaushik Roy Purdue University EMAIL |
| Pseudocode | No | The paper describes methods and architectures in text and diagrams (Figure 2, Figure 4) but does not present any formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | *Code is available at https://github.com/ar-roy/dct-cryptonets |
| Open Datasets | Yes | Each network is trained on CIFAR-10, mini-Image Net, Imagenette, and Image Net, with varying configurations of DCT components. |
| Dataset Splits | No | The paper mentions using well-known datasets like CIFAR-10 and Image Net, and refers to the 'full Image Net validation set'. However, it does not explicitly state specific percentages, sample counts, or detailed methodology for how the training, validation, and test splits were performed for any of the datasets. |
| Hardware Specification | Yes | Neural network training is conducted on an RTX A40 GPU and FHENN latency measurements are obtained from dual AMD Ryzen Threadripper PRO 5965WX processors (96-threads total). |
| Software Dependencies | No | The paper mentions 'libjpeg-turbo', 'Brevitas (Pappalardo, 2023)', and 'Concrete-ML (Zama, 2022)' as tools and frameworks used. However, specific version numbers for these software components are not explicitly provided in the text; the years in parentheses refer to publications or repository updates, not software versions. |
| Experiment Setup | Yes | Our approach employs a 4-bit symmetric quantization scheme with a fixed zero-point for all tasks except Image Net, where 5-bit quantization is used. These choices strike a balance between efficiency and accuracy, allowing us to perform the entire neural network computation directly on encrypted data within the TFHE framework. (...) Table 8: Training, Quantization and Cryptographic Hyperparameters. Section Hyperparameter CIFAR-10 mini-Image Net Imagenette Image Net Epochs 60 90 Batch Size 32 256 Optimizer Adam Adam Learning Rate 1e-3 1e-3 Weight Decay 1e-5 1e-5 Gradient Clipping 0.1 0.1 Dropout 0.2 0.2 Scheduler [20, 45] [30, 60] Decay Factor 0.1 0.1 Quantization Weight Bit-Width 4 4 Weight Quantization Protocol Int8Weight Per Tensor Float Int8Weight Per Tensor Float Activation Bit-Width 4 4 Activation Quantization Protocol Int8Act Per Tensor Float Int8Act Per Tensor Float Cryptographic Number of Bits 5 5 PBS Error Probability 0.01 0.01 Bit-removal Rounding Threshold 6 7 |