List your company
ResearchFormerly WizWand

Secret Data Extraction Attack on FedLLM 100 Target Secrets

100Success Rate (P > .10)GPT-2 Small (124M)
-4235077Jun 15, 2026
Updated 4mo ago

Evaluation results

MethodLinks
GPT-2 Small (124M)Setting=After poisoning2026.0610093
GPT-Neo (125M)Setting=After poisoning2026.0610094
Pythia (160M)Setting=After poisoning2026.0610096
OPT (250M)Setting=After poisoning2026.0610099
GPT-2 Medium (355M)Setting=After poisoning2026.06100100
Llama2 (13B)Setting=After poisoning2026.0610097
DistilGPT-2 (82M)Setting=After poisoning2026.069985
Llama2 (7B)Setting=After poisoning2026.069983
Llama3.2 (1B)Setting=After poisoning2026.069792
GPT-2 Medium (355M)Setting=Baseline setting2026.06237
OPT (250M)Setting=Baseline setting2026.06213
Pythia (160M)Setting=Baseline setting2026.06170
GPT-Neo (125M)Setting=Baseline setting2026.06130
GPT-2 Small (124M)Setting=Baseline setting2026.06120
DistilGPT-2 (82M)Setting=Baseline setting2026.0690
Llama2 (7B)Setting=Baseline setting2026.0600
Llama2 (13B)Setting=Baseline setting2026.0600
Llama3.2 (1B)Setting=Baseline setting2026.0600