Abstract
Privacy in video streaming is often compromised as Content Providers (CPs) construct detailed user profiles from request traces. This paper introduces PPVF, a privacy-preserving video fetching framework that utilizes trusted edge devices to camouflage user interests. By integrating Federated Learning for popularity prediction and Correlated Differential Privacy (CDP) for request generation, PPVF generates "smart noise"—redundant requests that are statistically similar to real user behavior. This approach protects privacy while simultaneously enhancing caching performance by pre-fetching potentially useful content. Experimental results on real-world datasets demonstrate that PPVF improves the Cache Hit Ratio by 18.15% and significantly degrades CP profiling capabilities.
Introduction
While transport-layer encryption (e.g., HTTPS) protects the content of video streams, the metadata of requests—specifically which videos are requested and when—remains visible to Content Providers. Over time, these traces allow CPs to infer sensitive user attributes such as age, gender, and personal interests. Existing privacy solutions, such as standard Differential Privacy (DP), often introduce random noise that clogs network bandwidth and fills caches with irrelevant data. PPVF addresses this by creating "smart noise": redundant requests that are realistic enough to fool profiling algorithms while remaining useful for local caching.
The PPVF Framework
The PPVF framework operates on trusted edge devices (EDs), such as smart routers or local CDN nodes. It consists of three primary pillars designed to balance privacy, utility, and resource constraints.
1. FL-based Video Utility Predictor
To ensure that redundant requests provide caching benefits, PPVF employs a Federated Learning (FL) framework. It utilizes a Mutual-Exciting Process (MEP) to model temporal and cross-video correlations. This allows the system to predict the "utility" of a video—the likelihood of a future request—collaboratively across edge devices without sharing raw user data.
2. Online Privacy Budget Scheduler
Privacy is managed as a finite resource via a privacy budget. The scheduler employs a threshold-based online algorithm to decide which "noise" videos to fetch. It prioritizes videos with high predicted utility and low privacy cost, ensuring the system stays within defined privacy bounds while maximizing cache efficiency.
3. CDP-based Request Generator
Unlike traditional DP, which treats data points as independent, PPVF uses Correlated Differential Privacy (CDP). By calculating the Pearson correlation between videos, the generator creates redundant requests that are statistically related to real requests. This makes the camouflage much harder for a CP to filter out using collaborative filtering techniques.
Experimental Evaluation
The framework was evaluated using the Tencent Video Traces dataset, which contains 933,541 viewing requests from 10,000 active users over a 30-day period. To support reproducibility, the authors have released the source code and a sampled version of this dataset.
Key Results
- Privacy Protection: PPVF reduced Jaccard Similarity (measuring the match between exposed and real profiles) by 17.54% to 22.38% compared to state-of-the-art privacy baselines.
- Caching Efficiency: Despite privacy constraints, PPVF achieved a Cache Hit Ratio (CHR) 18.15% higher than non-privacy-preserving models like HRS.
- Adversarial Resistance: When tested against a Neural Collaborative Filtering (NCF) recommendation engine acting as an adversary, PPVF successfully degraded the engine's ability to accurately profile users.
Conclusion
PPVF demonstrates that privacy and streaming performance can be synergistic. By leveraging edge computing and correlated noise, the framework provides a robust solution for shielding user habits. Future work will explore reducing computational overhead and adapting the framework for heterogeneous network environments with varying storage and processing capabilities.
Summary updated