MSG1999/DLOps-A5-Q2-ART

DLOps Assignment 5 — Q2: Adversarial Attacks with IBM ART

This repository contains three PyTorch model weights trained on CIFAR-10 as part of an adversarial robustness study using IBM ART.

Models

File	Architecture	Task	Best Accuracy
resnet18_cifar10_best.pt	ResNet-18 (CIFAR adapted)	10-class classification	94.68% val acc
detector_BIM_best.pt	ResNet-34 (CIFAR adapted)	BIM adversarial detector (binary)	99.57% detection acc
detector_PGD_best.pt	ResNet-34 (CIFAR adapted)	PGD adversarial detector (binary)	99.93% detection acc

Architecture Notes

• ResNet-18 classifier: 3×3 stem conv (stride 1), no maxpool, 10-class head
• ResNet-34 detectors: 3×3 stem conv (stride 1), no maxpool, 2-class head (clean=0, adversarial=1), internal CIFAR-10 normalization

FGSM Attack Results (Part i)

ε	FGSM-Scratch	FGSM-ART	Drop (Scratch)
0.01	48.70%	52.55%	45.55%
0.05	33.80%	35.90%	60.45%
0.10	16.80%	17.45%	77.45%
0.30	9.90%	9.95%	84.35%

Clean accuracy: 94.25%

Usage

import torch
import torch.nn as nn
from torchvision import models

def build_resnet18():
    m = models.resnet18(pretrained=False)
    m.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False)
    m.maxpool = nn.Identity()
    m.fc = nn.Linear(m.fc.in_features, 10)
    return m

model = build_resnet18()
state = torch.load("resnet18_cifar10_best.pt", map_location="cpu")
model.load_state_dict(state)
model.eval()

WandB Logs

DLOps-A5-Q2-ART on W&B