Density Feedback Control for Transferability of Generative Sparse Adversarial Attacks
Organizations: Korea Aerospace University Goyang, Republic of Korea
Abstract
Generator-based sparse adversarial attacks train encoder--decoder networks to produce transferable sparse perturbations for black-box robustness evaluation. Their perturbation density, however, is set only indirectly by a fixed sparsity coefficient whose mapping to density is unknown and shifts during training, so reaching a target density requires re-tuning; this fixed pressure also yields little transferability gain once density reduction begins. We propose Density Feedback Control (DFC), which tightens or relaxes this coefficient at the end of each epoch according to the gap between the observed hard-mask density and the target, without changing the generator architecture or loss. At a 10% target density, DFC brings the mean density of three base attacks within one percentage point of the target in a single run and improves the transfer attack success rate (ASR) of two of them on CNN and Transformer targets, including under per-image budgets, while perturbing fewer pixels. Tracing GPG's training, we observe that as DFC relaxes the coefficient, the transfer ASR recovers with the density and overtakes the fixed-coefficient baseline with fewer pixels. For EGS-TSSA, which restricts perturbations to source-model feature-importance regions, DFC yields no consistent gain; we analyze how this restriction interacts with density feedback.
Figures & tables
| Method | Density | ResNet-50 | IncV3 | VGG16 | Dense161 | WideRes50 | EffNetB6 | Mean ASR |
|---|---|---|---|---|---|---|---|---|
| PGD 0 ( ) | 10.00 | 100.00 ∗ | 10.09 | 20.94 | 23.54 | 29.27 | 10.09 | 18.78 |
| sPGD proj ( ) | 9.99 | 99.95 ∗ | 6.93 | 16.25 | 18.35 | 22.59 | 8.16 | 14.46 |
| sPGD unproj ( ) | 9.99 | 99.95 ∗ | 9.93 | 21.98 | 26.08 | 32.99 | 10.39 | 20.27 |
| SparseFool ( ) | 12.90 | 100.00 ∗ | 9.04 | 15.83 | 15.53 | 18.24 | 9.74 | 13.67 |
| GreedyFool ( ) | 12.72 | 100.00 ∗ | 9.01 | 14.95 | 19.83 | 24.40 | 9.80 | 15.60 |
| GPG | 11.69 | 80.67 ∗ | 18.44 | 42.49 | 35.80 | 54.73 | 13.50 | 32.99 |
| Method | Density | ResNet-50 | ViT-S | DeiT-S | TNT-S | Swin-T | Twins-S | Mean ASR |
|---|---|---|---|---|---|---|---|---|
| PGD 0 ( ) | 10.00 | 100.00 ∗ | 6.73 | 4.37 | 7.97 | 7.34 | 7.83 | 6.85 |
| sPGD proj ( ) | 9.99 | 99.95 ∗ | 5.69 | 3.67 | 6.45 | 6.45 | 5.99 | 5.65 |
| sPGD unproj ( ) | 9.99 | 99.95 ∗ | 6.49 | 4.69 | 7.76 | 7.57 | 8.13 | 6.93 |
| SparseFool ( ) | 12.90 | 100.00 ∗ | 5.77 | 2.72 | 5.54 | 4.07 | 4.82 | 4.58 |
| GreedyFool ( ) | 12.72 | 100.00 ∗ | 6.73 | 3.74 | 6.19 | 6.20 | 6.69 | 5.91 |
| GPG | 11.69 | 80.67 ∗ | 8.78 | 5.60 | 10.49 | 9.43 | 11.12 | 9.08 |
| Method | |||
|---|---|---|---|
| PGD 0 | 12.82 | 13.12 | 13.65 |
| sPGD proj | 10.06 | 10.36 | 10.75 |
| sPGD unproj | 13.60 | 14.25 | 14.81 |
| GPG | 7.95 | 10.81 | 13.51 |
| GPG-DFC | 13.74 (+72.68%) | 16.82 (+55.61%) | 19.63 (+45.30%) |
| TSAA | 6.41 | 8.94 | 11.41 |
| Method | Total | Density | Success | : B D | : B D |
|---|---|---|---|---|---|
| GPG | +5.78 | +2.60 | +3.19 | 20.22 29.84 | 21.53 27.28 |
| TSAA | +9.31 | +3.79 | +5.52 | 18.24 33.59 | 12.99 19.30 |
| EGS-TSSA | 0.53 | 0.51 | 0.01 | 24.31 17.79 | 13.13 17.82 |
| ASR change (pp) | ||||
| Base | (%) | Total | Dens. | Succ. |
| + GFI mask | ||||
| TSAA | 10.86 10.22 | |||
| TSAA-DFC | 0 9.99 10.89 | |||
| + DFC | ||||
| TSAA | 10.86 0 9.99 | |||
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
| Run | Excess | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Baseline | 6 | 18.01 | 5.89 | 6.72 | 34.29 | 24.52 | 7.31 | 1.91 | 5.39 |
| 7 | 17.25 | 6.67 | 5.91 | 37.06 | 24.25 | 8.38 | 2.18 | 6.20 | |
| 8 | 16.51 | 6.74 | 6.00 | 39.09 | 23.35 | 8.58 | 2.12 | 6.45 | |
| 9 | 17.47 | 5.11 | 6.07 | 30.96 | 26.65 | 6.31 | 1.57 | 4.73 | |
| 10 | 16.43 | 6.41 | 5.36 | 36.67 | 25.11 | 8.24 | 1.84 | 6.40 | |
| 11 | 15.26 | 6.10 | 4.94 | 37.14 | 23.22 | 8.63 | 1.77 | 6.86 |
| Method | Without | With | Drop (pp) |
|---|---|---|---|
| EGS-TSSA | 22.44 | 17.19 | 5.25 |
| EGS-TSSA-DFC | 25.36 | 16.74 | 8.62 |
| Baseline | DFC | |||
|---|---|---|---|---|
| Method | Red./Stab. | Fine-tune | ||
| GPG | ||||
| TSAA | ||||
| EGS-TSSA | ||||
Adversarial Attacks
Adversarial attacks exploit crafted inputs, poisoned data, or system interactions to manipulate AI predictions, behavior, privacy, or reliability.
See all 477 papers