Research map / Architectures and efficiency / Attention and transformers
SAL-ViT: Towards Latency Efficient Private Inference on ViT using Selective Attention Search with a Learnable Softmax Approximation
Yuke Zhang, Dake Chen, Souvik Kundu, Chenghao Li, Peter A. Beerel
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use external attention instead of self-attention | ViT on CIFAR-10 · CIFAR-10 | approx PI latency (s): 81.48 -> 41.72 | Improved |
| use external attention instead of self-attention | ViT on CIFAR-10 · CIFAR-10 | accuracy (%): 95.56 -> 92.86 | Hurt |
| replace softmax with scaling function in EA | EA ViT · CIFAR-10 | accuracy (%): 92.86 -> 33.45 | Hurt |
| replace SA softmax with 2Quad | SA ViT · CIFAR-10 | approx PI latency (s): 81.48 -> 51.12 | Improved |
| make L2Q parameters finer-grained (element-wise vs global) (ablation) | L2Q · CIFAR-100 | accuracy (%): 76.08 -> 78.25 | Improved |
| make both shifting and scaling parameters learnable (ablation) | element-wise L2Q · CIFAR-100 | accuracy (%): 77.6 -> 78.25 | Improved |
| select SA layers by learned alpha (SAS-L) rather than lowest alpha (worst case) (ablation) | SAS · CIFAR-10/CIFAR-100/Tiny-ImageNet | accuracy (%) | Improved |
| use head-wise search (SAS-H) instead of layer-wise (SAS-L) (ablation) | SAS · CIFAR-100 | accuracy (%): 77.62 -> 77.15 | Hurt |
| use 2ReLU in SAS-L (ablation) | SAS-L · Tiny-ImageNet | accuracy (%) | Hurt |
| apply SAL-ViT hybrid attention with L2Q instead of full SA softmax | CCT ViT · ImageNet | accuracy (%): 77.22 -> 75.51 accuracy, 728.22 -> 450.87 s latency | Mixed |
Closest papers by meaning
- DCT-CryptoNets: Scaling Private Inference in the Frequency Domain (ICLR 2025 · similarity 0.86)
- DeepCompress-ViT: Rethinking Model Compression to Enhance Efficiency of Vision Transformers at the Edge (CVPR 2025 · similarity 0.85)
- Jaguar: Fast Private CNN Inference with Power-of-Two Homomorphic Arithmetic (NeurIPS 2026 · similarity 0.85)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.85)
- ViP: A Differentially Private Foundation Model for Computer Vision (ICML 2024 · similarity 0.84)
- CipherPrune: Efficient and Scalable Private Transformer Inference (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
