List your company
ResearchFormerly WizWand

Attacker Detection on Qwen3-1.7B target τpool=1.509 (test)

53.9Tau Multiplier (×τ)OLMo-2-7B + public φ
-1.11613.16727.4541.733Apr 20, 2026
Updated 6mo ago

Evaluation results

MethodLinks
OLMo-2-7B + public φClass=cross, Budget=E32026.0453.981.3———
Qwen2.5-1.5B + public φClass=cross, Budget=E32026.0445.969.2———
Qwen3-0.6B + public φClass=cross, Budget=E32026.0441.662.8———
library-aware φ onlyClass=LoRA, Budget=E4-v2 baseline2026.043857.3821.2——
E-E α=0.1, λ=0Class=whitebox, Budget=r=64, 3000 steps2026.043857.38———
E-E α=1, λ=0.1 (seed 2)Class=whitebox, Budget=r=64, 3000 steps2026.043857.383,591——
E-E α=1, λ=0Class=whitebox, Budget=r=64, 3000 steps2026.0437.556.55———
Phi-3.5-mini + public φClass=cross, Budget=E32026.0434.752.4———
E-E α=0, λ=0.1Class=whitebox, Budget=r=64, 3000 steps2026.0432.849.463,625——
r=128 MSE, λ=0.1Class=LoRA, Budget=r=128, 3000 steps2026.0431.147.013,538——
public-φ + LoRA r=16Class=LoRA, Budget=3000 steps2026.041218.125——
E-E α=0.1, λ=0.1Class=whitebox, Budget=r=64, 3000 steps2026.046.629.991,188——
E-E α=1, λ=0 (seed 2)Class=whitebox, Budget=r=64, 3000 steps2026.046.449.7186,809——
E4-v2 STAGEB λ=1.0Class=LoRA, Budget=r=64, 1500 steps2026.044.967.4850.98——
E-E α=1, λ=0.1Class=whitebox, Budget=r=64, 3000 steps2026.044.97.4243——
E-E αjz=0, λ=0Class=whitebox, Budget=r=64, 3000 st., 3 seeds†2026.044.576.8931.9——
r=128 MSE, λ=0Class=LoRA, Budget=r=128, 3000 steps2026.044.256.4250.37——
E4-v2 STAGEC λ=0.1Class=LoRA, Budget=r=64, 1500 steps2026.044.166.2763.16——
E4-v2 STAGEA λ=0Class=LoRA, Budget=r=64, 1500 steps2026.042.784.1932.84——
ZERO_PAD Qwen3-0.6BClass=lift, Budget=E2, natt=482026.041————
RAND_PROJ Qwen3-0.6BClass=lift, Budget=E2, natt=482026.041————
LEARNED_LIN Qwen3-0.6BClass=lift, Budget=E2, natt=482026.041————