Website fingerprinting data format (WFDF): Harnessing parallelisation performance on dataset encoding and decoding operations

Website Fingerprinting (WF) technique analyses encrypted network communication traffic to identify webpages visited by Tor users. State-of-the-art WF techniques rely on deep learning (DL) models to extract complex traffic patterns. However, the massive dataset requirements of DL create severe input/output (I/O) bottlenecks. While existing research focuses on optimising the computational mathematics of model training, the dataset loading process is largely overlooked. Because researchers must iteratively construct DL models many times, slow dataset loading creates a severe compounding delay in the experimental lifecycle. To improve the dataset handling process, this study proposes the Website Fingerprinting Dataset Format (WFDF), a specialised binary file format designed for WF datasets. WFDF replaces existing sequential I/O with a multi-threaded, chunk-based encoding and decoding algorithm that processes massive multi-dimensional arrays in parallel. It natively integrates zero-copy memory mapping, Zstandard compression, and cryptographic hashing to ensure data integrity without imposing computational penalties. Our evaluations demonstrate that WFDF achieves better data compression, reducing storage footprints to an average of 21% of their raw dataset size. Furthermore, WFDF delivers better execution performance, recording the fastest mean load and write times of 4.91 s and 17.99 s, with maximum standard deviations of 0.08 s and 0.48 s, respectively. This yields a 5.04 × to 6.59 × loading speedup compared to popular existing formats. A Friedman test confirms that the overall performance improvements across the evaluated formats are statistically significant (p < 0.001). By reducing storage burdens and latency, WFDF provides a scalable solution that accelerates the experimental lifecycle of DL-based WF models.