DeepSFU: Scalable Deepfake Detection for Video Conferencing
Published in ACM SIGCOMM, 2026
Recommended citation: T. Tran, S. Ebadi, S. M. H. Hosseini, W. Shin, E. Ram, Y. Son, S. Kim, N. Bui, K. Lee, E. Keller, S. Ha. "DeepSFU: Scalable Deepfake Detection for Video Conferencing." ACM SIGCOMM, 2026. https://dl.acm.org/doi/10.1145/3789240.3829184
Real-time deepfake detection as a native SFU service: detects deepfake onsets directly on the encoded VP9 bitstream and offloads crypto and NIC-to-GPU frame movement to a BlueField-3 DPU — ~144× lower per-frame processing latency and 26× more concurrent conferences at 95.8% precision.