A Multi-FPGA System for End-to-End Offline Password Recovery Acceleration (www.mdpi.com)

Researchers introduced PassRecover, a multi-FPGA system (one XCZU7EV + two XCVU9P) that for the first time accelerates an end-to-end offline password recovery pipeline by co‑accelerating a deep learning password generator (PassGAN) and the target encryption/verification algorithms on the same reconfigurable hardware. Their custom NPU for PassGAN (Conv1D-based GAN generator with tiling) achieves a reported 82.16% higher throughput and 155.22% better energy efficiency than a Tesla V100 GPU. Complementary encryption accelerators—implemented from a unified hardware template and tailored for Office 2010/2013, PDF 1.7, WINZIP and RAR5—are integrated into the pipeline; compared to the latest “encryption-only” FPGA work, the end‑to‑end design delivers on average 101.5% higher speed and 22.11% better energy efficiency. Technically, the design combines a tiled NPU datapath optimized for Conv1D layers used by PassGAN with specialized encryption threads that handle compute-heavy primitives (e.g., SHA-512 loops with 100k iterations, dynamic padding, and AES‑256‑CBC operations for Office2013 verification). The co‑design balances generator throughput and crypto verification to avoid bottlenecks, demonstrates resource usage across the FPGAs (LUT utilization ~60%), and evaluates scalability and reconfiguration. The work shows that tightly integrated FPGA pipelines can outperform GPUs on GAN-based password generation plus real-world crypto checks—raising both practical capabilities for defenders/offense and important considerations for password security.