|  | Budget 10 | Budget 30 | Budget 50 |
---|
Defense | Training | Net | FGA | IG | Net | FGA | IG | Net | FGA | IG |
---|
Jaccard | Rand. | 0.224 | N/A | N/A | 0.576 | N/A | N/A | 0.744 | N/A | N/A |
Jaccard | SD | 0.12 | N/A | N/A | 0.136 | N/A | N/A | 0.16 | N/A | N/A |
Jaccard | GC | 0.128 | N/A | N/A | 0.192 | N/A | N/A | 0.2 | N/A | N/A |
GCN | Rand. | 0.456 | N/A | N/A | 0.76 | N/A | N/A | 0.888 | N/A | N/A |
GCN | SD | 0.6 | N/A | N/A | 0.936 | N/A | N/A | 0.976 | N/A | N/A |
GCN | GC | 0.544 | N/A | N/A | 0.88 | N/A | N/A | 0.952 | N/A | N/A |
Cheb | Rand. | 0.056 | N/A | N/A | 0.072 | N/A | N/A | 0.072 | N/A | N/A |
Cheb | SD | 0.072 | N/A | N/A | 0.128 | N/A | N/A | 0.136 | N/A | N/A |
Cheb | GC | 0.136 | N/A | N/A | 0.16 | N/A | N/A | 0.192 | N/A | N/A |
- Results are included for Nettack (Net), FGA, and IG-FGSM (IG). For each classifier, we train with random (Rand.), StratDegree (SD), and GreedyCover (GC). Each entry is a probability of attack success, thus higher is better for the attacker and lower is better for the defender. To yield the most robust classifier, the defender picks the classifier/training method combination that minimizes the worst-case attack probability. These entries are listed in bold. Entries listed as N/A did not finish in the allotted time (24Â h per trial). Only results using Jaccard, GCN, and ChebNet were obtained in time. While StratDegree and GreedyCover improve performance with the Jaccard-based classifier, the best performance is achieved by a ChebNet classifier with random training. In our experiments, this classifier with the PubMed data typically has a much higher margin before the attack takes place