PrivAware

# PRISM-DECODE: Privacy-Preserving Scene Understanding for Public Health Surveillance ## Overview This repository implements **PRISM (Privacy-Regularized Inference and Structure Modeling)** and **DECODE (Decentralized Compression and Obfuscation for Differential Exposure)** — a novel framework for **privacy-preserving AI-driven public health surveillance**. Our approach integrates privacy constraints directly into model architecture and training workflows to mitigate privacy leakage while maintaining strong scene understanding performance. ## Motivation AI-based public health surveillance systems often analyze sensitive visual data from public environments, raising concerns about privacy, consent, and ethical governance. Traditional methods (e.g., anonymization, obfuscation) are insufficient against modern threats such as:- Membership inference attacks- Attribute inference- Latent feature reconstruction PRISM-DECODE addresses these threats by embedding privacy preservation into the AI model’s computational structure and training dynamics. ## Key Contributions - **PRISM Framework**: - Decomposes latent space into utility- and sensitive-related subspaces. - Uses adversarial training and orthogonal constraints to disentangle sensitive information. - Suppresses mutual information between latent representations and sensitive attributes. - **DECODE Protocol**: - Dynamically regulates gradient and feature exposure using entropy-aware dropout and perturbation. - Enforces distributional alignment with a reference prior and constrains batch-wise statistical leakage. - Promotes invariance under sensitive attribute perturbations via auxiliary encoders. - **Sensitivity-Aware Modules**: - Estimate and suppress latent unit responsiveness to protected attributes. - Apply selective attention masking at the feature level. ## Architecture - **Encoder-Decoder Backbone** with U-Net-style skip connections.- **Dilformer Blocks**: Multi-scale dilated convolutions for high-resolution feature extraction.- **Adversarial Discriminator**: Attempts to decode sensitive attributes; used to train obfuscation.- **Entropy-based Regularizers**: Applied to both feature magnitude and training gradients. (Refer to *Figure 1–4 in the paper* for detailed architectural diagrams.) ## Datasets Evaluated across 4 challenging public datasets:- **Masked Faces Dataset**: Occlusion-aware facial identity recognition.- **Avenue Dataset**: Anomaly detection in surveillance videos.- **DukeMTMC**: Multi-camera person re-identification.- **Market-1501 / MSMT17**: Benchmark person re-ID with occlusions and lighting variation. ## Results PRISM-DECODE demonstrates state-of-the-art performance in both utility and privacy metrics: | Dataset | Accuracy | F1 Score | AUC ||--------------------|----------|----------|---------|| Masked Faces | 93.42% | 91.08% | 94.10% || Avenue | 92.27% | 89.88% | 93.76% || DukeMTMC | 93.11% | 90.94% | 94.28% || RE-IDENTIFICATION | 96.20% | 93.90% | 95.71% | Compared to leading baselines (CLIP, ViT, BLIP, T5), our model shows consistent improvement, especially in privacy-sensitive benchmarks. ## How It Works The loss function integrates multiple objectives: ```pythonL_total = L_pred + L_priv + λ_align * L_align + λ_inv * L_inv + λ_env * ExposurePenalty# PRISM-DECODE：面向公共卫生监测的隐私保护场景理解 ## 概述 本开源仓库实现了**PRISM（隐私正则化推理与结构建模，Privacy-Regularized Inference and Structure Modeling）**与**DECODE（差分暴露去中心化压缩与混淆，Decentralized Compression and Obfuscation for Differential Exposure）**——一款面向隐私保护的AI驱动公共卫生监测新型框架。我们的方法将隐私约束直接嵌入模型架构与训练流程，在保障场景理解性能的同时，有效缓解隐私泄露风险。 ## 研究动机 基于AI的公共卫生监测系统通常会分析公共环境中的敏感视觉数据，由此引发隐私、知情同意与伦理治理层面的诸多争议与担忧。传统方法（如匿名化、混淆处理）难以抵御现代攻击威胁，具体包括：成员推理攻击、属性推理攻击、潜在特征重构。PRISM-DECODE通过将隐私保护机制嵌入AI模型的计算结构与训练动态，针对性解决上述安全隐患。 ## 核心贡献 - **PRISM框架**： - 将潜在空间拆解为效用相关子空间与敏感相关子空间 - 采用对抗训练与正交约束实现敏感信息的解耦 - 抑制潜在表征与敏感属性之间的互信息 - **DECODE协议**： - 基于感知熵的Dropout与扰动策略，动态调节梯度与特征暴露程度 - 强制模型分布与参考先验对齐，约束批次级统计信息泄露 - 通过辅助编码器实现敏感属性扰动下的特征不变性 - **敏感感知模块**： - 估计并抑制潜在单元对受保护属性的响应性 - 在特征层面应用选择性注意力掩码 ## 架构设计 - 搭载U-Net风格跳跃连接的**编码器-解码器骨干网络** - **Dilformer模块**：用于高分辨率特征提取的多尺度空洞卷积结构 - **对抗判别器**：尝试解码敏感属性，用于训练混淆机制 - **基于熵的正则器**：同时应用于特征幅值与训练梯度 （详细架构图请参见论文中的图1至图4。） ## 测试数据集 本研究在4个具有挑战性的公开数据集上开展评估： - **遮罩人脸数据集（Masked Faces Dataset）**：面向遮挡感知的人脸身份识别任务 - **Avenue数据集（Avenue Dataset）**：监控视频异常检测任务 - **DukeMTMC数据集（DukeMTMC）**：多摄像头行人重识别任务 - **Market-1501 / MSMT17**：包含遮挡与光照变化的行人重识别基准数据集 ## 实验结果 PRISM-DECODE在效用指标与隐私指标上均展现出当前最优性能： | 数据集 | 准确率 | F1分数 | AUC值 | |---------------------|---------|---------|---------| | 遮罩人脸 | 93.42% | 91.08% | 94.10% | | Avenue数据集 | 92.27% | 89.88% | 93.76% | | DukeMTMC数据集 | 93.11% | 90.94% | 94.28% | | 行人重识别 | 96.20% | 93.90% | 95.71% | 与主流基线模型（CLIP、ViT、BLIP、T5）相比，本模型展现出一致的性能提升，尤其在隐私敏感基准测试中表现突出。 ## 工作原理 损失函数整合了多项目标： python L_total = L_pred + L_priv + λ_align * L_align + λ_inv * L_inv + λ_env * ExposurePenalty