The authors design a two-stage attack.
Backdoor attack on federated prompt tuning exploits frozen backbone to evade detection
ShadowFPT uses a learnable shadow trigger to steer triggered inputs toward target class in representation space, keeping prompt updates benign-appearing.
Chinese Tech
Kun Zhai · Teng Li · Yunhao Feng · Xingjun Ma
Fudan University · Alibaba Group
Research Digest··2 min read
The authors propose ShadowFPT, a backdoor attack targeting federated prompt tuning (FPT) that exploits the shared frozen VLM backbone rather than corrupting prompt updates.
§