RTF-Net: A Lightweight Dual-View Framework for Fine-Grained Encrypted Short-Video Fingerprinting

Short video platforms like TikTok have dominated global internet traffic. Despite the deployment of robust encryption protocols (e.g., TLS 1.3 and QUIC) to protect user confidentiality, side-channel information leaking from encrypted traffic remains a significant privacy threat. However, accurately identifying specific video content is increasingly challenging due to complex Adaptive Bitrate (ABR) algorithms, which homogenize traffic patterns, especially in high-motion viral content. Existing methods relying on simple statistical features or single-view deep learning models often struggle to capture the fine-grained fingerprints of specific videos under these conditions. In this paper, we propose a novel Multi-View Fusion Network, namely Rocket-Transformer Fusion (RTF-Net), that integrates time-domain convolutional features with frequency-domain attention mechanisms. Our approach utilizes MiniRocket to extract deterministic sequential patterns from traffic bursts, while simultaneously employing Fast Fourier Transform (FFT) combined with TinyBERT to capture global periodic dependencies in the frequency domain. We evaluate our method on a self-collected dataset comprising 2,000 encrypted traffic traces generated from 20 most viral short dramas (each exceeding 100 million views) on TikTok. Experimental results demonstrate that our method achieves a state-of-the-art classification accuracy of 91.33% using only the first 30 seconds of viewing traffic, significantly outperforming other baseline machine learning algorithms, showing that side-channel leakage is indeed a threat to short video platforms. To facilitate reproducibility and future research, we have open-sourced our dataset and code online.