Multi-Dataset Backdoor Mitigation in Model Merging (CIFAR10, Cars196, SVHN, SUN397, STL10, Pets)
95.61CIFAR10 CAIndividual
Evaluation Results
| Method | Links | |||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| IndividualAttack Type=LoBAM, Task Sequence=Task Order 2, Adversary Task=CIFAR10, Target Task=Cars1962026.06 | 95.61 | 67.88 | 61.65 | 54.33 | 93.29 | 17.67 | 67.93 | 48.38 | 97.32 | 2.51 | 89.56 | 18.28 | 84.22 | 54.33 | 30.94 | |
| SAMAttack Type=LoBAM, Task Sequence=Task Order 2, Adversary Task=CIFAR10, Target Task=Cars1962026.06 | 86.84 | 12.24 | 50 | 7.22 | 23.67 | 70.48 | 62.21 | 30.5 | 95.96 | 1.78 | 86.12 | 15.72 | 67.46 | 7.22 | 26.14 | |
| LFPMAttack Type=LoBAM, Task Sequence=Task Order 2, Adversary Task=CIFAR10, Target Task=Cars1962026.06 | 86.79 | 5.47 | 47.99 | 2.94 | 20.51 | 26.88 | 61.88 | 23.22 | 96.45 | 1.32 | 84.98 | 9.34 | 66.43 | 2.94 | 13.24 | |
| PAMAttack Type=LoBAM, Task Sequence=Task Order 2, Adversary Task=CIFAR10, Target Task=Cars1962026.06 | 85.59 | 13.21 | 48.24 | 8.36 | 26.35 | 64.33 | 61.36 | 33.42 | 95.46 | 2.17 | 85.55 | 18.2 | 67.09 | 8.36 | 26.26 | |
| TAAttack Type=LoBAM, Task Sequence=Task Order 2, Adversary Task=CIFAR10, Target Task=Cars1962026.06 | 84.42 | 37.05 | 47.45 | 92.66 | 26.9 | 97.21 | 57.2 | 90.32 | 95.07 | 13.16 | 81.57 | 36.22 | 65.43 | 92.66 | 54.79 | |
| SAUAttack Type=LoBAM, Task Sequence=Task Order 2, Adversary Task=CIFAR10, Target Task=Cars1962026.06 | 81.62 | 37.89 | 40.25 | 22.98 | 30.03 | 80.45 | 57.2 | 67.68 | 94.45 | 20.73 | 83.21 | 29.19 | 64.46 | 22.98 | 47.18 | |
| IBVSAttack Type=LoBAM, Task Sequence=Task Order 2, Adversary Task=CIFAR10, Target Task=Cars1962026.06 | 81.37 | 25.38 | 46.73 | 19.87 | 21.17 | 53.86 | 53.52 | 61.73 | 94.28 | 8.7 | 76.99 | 34.5 | 62.34 | 19.87 | 36.83 |