Groundbreaking Results: Setting New Benchmarks

Our research has achieved unprecedented levels of robustness against AutoAttack, the most rigorous ensemble of adversarial attacks, particularly on a critical medical dataset:

BrainTumor-7K (Medical Imaging Dataset - ResNet-18 Model):

• Baseline AutoAttack Accuracy (ϵ=0.020): 48.29%

• EDF Evolutionary Model AutoAttack Accuracy (ϵ=0.020): 98.86%

• EDF Evolutionary Model AutoAttack Accuracy (ϵ=0.050): 97.86%

• Improvement: An astonishing 50.57% gain in AutoAttack robustness over the baseline.

This achievement of nearly 99% AutoAttack robustness on a medical imaging dataset is, to our knowledge, a new state-of-the-art benchmark, profoundly impacting the potential for reliable AI deployment in healthcare.

Additional Highlights:

• Phishing Legitimate (Tabular Cybersecurity): Achieved 94.10% AutoAttack Accuracy (from 73.25% baseline) – potentially the highest reported for tabular data without adversarial training.

• Overall PGD Robustness (ϵ=0.020): Achieved a 75.78% improvement over baseline.

• Maintained Clean Accuracy: A negligible -0.14% change (from 0.9872 to 0.9858), showcasing the true elimination of the trade-off.