EvalDNN

A Toolbox for Deep Neural Network Models

Accuracy

* Official reported data is put in parentheses

Model Top-1 Top-5
densenet121 75.0%(75.0%) 92.3%(92.3%)
densenet169 76.2%(76.2%) 93.2%(93.2%)
densenet201 77.3%(77.3%) 93.6%(93.6%)
inception_resnet_v2 80.4%(80.3%) 95.3%(95.3%)
inception_v3 77.9%(77.9%) 93.8%(93.7%)
mobilenet 70.3%(70.4%) 89.5%(89.5%)
mobilenet_v2 71.2%(71.3%) 90.3%(90.1%)
nasnet_large 82.7%(82.5%) 96.2%(96.0%)
nasnet_mobile 73.8%(74.4%) 91.5%(91.9%)
resnet101 76.4%(76.4%) 92.8%(92.8%)
resnet101_v2 76.9%(77.2%) 93.7%(93.8%)
resnet152 76.6%(76.6%) 93.1%(93.1%)
resnet152_v2 77.7%(78.0%) 94.0%(94.2%)
resnet50 74.9%(74.9%) 92.1%(92.1%)
resnet50_v2 75.3%(76.0%) 92.9%(93.0%)
vgg16 71.3%(71.3%) 90.0%(90.1%)
vgg19 71.3%(71.3%) 90.0%(90.0%)
xception 79.0%(79.0%) 94.5%(94.5%)

Neuron Coverage

Model Layers Neurons t=0.0 t=0.1 t=0.2 t=0.3 t=0.4 t=0.5 t=0.6 t=0.7 t=0.8 t=0.9
densenet121 428 130315 98.8% 94.0% 86.4% 78.8% 72.8% 69.2% 66.7% 39.7% 7.1% 1.4%
densenet169 596 243979 98.0% 91.9% 86.0% 79.1% 72.9% 69.0% 66.5% 38.2% 6.5% 0.9%
densenet201 708 350987 97.5% 90.5% 83.8% 77.0% 71.3% 68.2% 55.0% 32.5% 5.0% 0.7%
inception_resnet_v2 781 247336 99.2% 96.4% 86.1% 75.5% 62.1% 51.0% 43.2% 28.6% 11.6% 3.7%
inception_v3 312 76264 100.0% 97.5% 87.3% 76.0% 63.5% 54.1% 46.6% 28.5% 11.3% 4.4%
mobilenet 92 39867 99.9% 99.6% 95.3% 88.9% 83.7% 78.5% 69.1% 49.5% 28.7% 19.3%
mobilenet_v2 156 53747 100.0% 98.0% 90.8% 86.1% 81.1% 76.9% 72.6% 55.8% 26.0% 6.7%
nasnet_large 1040 503404 100.0% 89.3% 72.6% 66.5% 63.9% 62.8% 59.3% 37.0% 10.4% 1.3%
nasnet_mobile 770 95718 99.9% 98.0% 87.2% 76.9% 69.1% 65.3% 62.7% 53.0% 27.8% 8.9%
resnet101 346 189867 100.0% 99.3% 91.6% 83.8% 78.5% 67.8% 58.0% 28.5% 4.9% 1.7%
resnet101_v2 378 195947 99.9% 89.6% 78.5% 73.8% 71.9% 71.1% 66.9% 28.7% 4.7% 1.4%
resnet152 516 274347 100.0% 99.0% 90.5% 82.7% 76.0% 66.5% 58.3% 29.6% 4.0% 1.2%
resnet152_v2 565 284267 99.9% 89.0% 77.6% 73.1% 71.4% 70.1% 63.9% 26.0% 3.5% 1.1%
resnet50 176 94123 100.0% 99.1% 93.3% 86.7% 81.2% 75.7% 65.8% 28.9% 6.6% 3.3%
resnet50_v2 191 95851 100.0% 92.7% 81.6% 75.8% 73.4% 72.2% 63.8% 30.0% 5.4% 2.8%
vgg16 21 14888 99.9% 96.5% 81.8% 71.5% 66.1% 63.4% 62.3% 61.9% 61.8% 61.7%
vgg19 24 16168 99.9% 95.7% 78.4% 67.1% 61.1% 58.5% 57.4% 57.0% 56.9% 56.8%
xception 133 91776 100.0% 97.6% 90.2% 83.7% 77.7% 74.8% 71.8% 40.9% 9.4% 3.5%

Robustness

Model FGSM BIM DeepFool
Success Rate Avg Time Avg Linf Dist Success Rate Avg Time Avg Linf Dist Success Rate Avg Time Avg MSE
densenet121 100.0% 0.24s 0.00006947 100.0% 4.65s 0.00000345 99.2% 2.27s 0.00000000
densenet169 100.0% 0.36s 0.00007872 100.0% 6.81s 0.00000357 99.4% 3.18s 0.00000000
densenet201 100.0% 0.46s 0.00008579 100.0% 9.10s 0.00000449 99.7% 3.52s 0.00000000
inception_resnet_v2 100.0% 9.40s 0.10819111 100.0% 32.51s 0.00615755 99.3% 6.33s 0.00001016
inception_v3 100.0% 2.06s 0.04405568 100.0% 12.70s 0.00144122 98.5% 3.30s 0.00000060
mobilenet 100.0% 0.15s 0.00164335 100.0% 2.65s 0.00062170 99.2% 0.62s 0.00000033
mobilenet_v2 100.0% 0.29s 0.00733108 100.0% 3.63s 0.00102919 99.3% 0.77s 0.00000042
nasnet_large 100.0% 20.07s 0.14748594 100.0% 59.18s 0.00648448 98.4% 14.49s 0.00000887
nasnet_mobile 100.0% 1.61s 0.03820292 100.0% 11.44s 0.00192647 98.7% 2.79s 0.00000144
resnet101 100.0% 1.30s 0.01351826 100.0% 15.78s 0.00141820 99.2% 3.86s 0.00000106
resnet101_v2 100.0% 2.28s 0.02718684 100.0% 20.01s 0.00226067 98.5% 5.74s 0.00000104
resnet152 100.0% 1.77s 0.01176379 100.0% 22.76s 0.00142831 99.2% 5.41s 0.00000105
resnet152_v2 100.0% 3.53s 0.03023599 100.0% 29.35s 0.00323015 97.9% 9.58s 0.00000114
resnet50 100.0% 0.58s 0.00725503 100.0% 8.40s 0.00117970 98.8% 2.31s 0.00000098
resnet50_v2 100.0% 1.04s 0.01932864 100.0% 10.89s 0.00184082 98.3% 3.08s 0.00000082
vgg16 100.0% 0.71s 0.00520756 100.0% 10.71s 0.00199782 99.7% 1.77s 0.00000256
vgg19 100.0% 0.90s 0.00598348 100.0% 13.33s 0.00224028 99.7% 2.17s 0.00000272
xception 99.9% 1.85s 0.04236282 100.0% 12.70s 0.00171585 98.9% 2.82s 0.00000102