nimaeb/xai-attack-detection-imagenette

XAI Attack Detection: Imagenette targeted BIM/PGD on ViT-B/16

Private research dataset of paired clean and targeted adversarial Imagenette images. It is built to study how adversarial attacks change a Vision Transformer's explanation maps and to support later work on attack detection. Each row is one source image with its clean and its attacked version.

Summary

Pairs	12,420 (train 8,690 · validation 1,860 · test 1,870)
Source images	Imagenette v2, 320 px (fastai/imagenette), resized 248 bicubic + center crop 224
Classifier	timm/vit_base_patch16_224.augreg2_in21k_ft_in1k at revision 063c6c38a5d8, 10 Imagenette logits taken from the ImageNet head, no fine-tuning
Attacks	targeted BIM, i.e. iterative FGSM (ε ∈ {2/255, 4/255, 8/255}, 10 steps, step size ε/10, no random start) and targeted L∞ PGD (ε ∈ {2/255, 4/255, 8/255}, steps ∈ {10, 20, 40}, step size 2.5·ε/steps, random start)
Target hit rate (after PNG)	88.9%
Label flip rate (after PNG)	88.9%
Plan signature	089f5ba81783f0cd

How it was built

1. The pinned ViT-B/16 classified all 13,394 Imagenette images with the 10 Imagenette logits of its 1000-class head. Fine-tuning was not needed (see notebook 01_model_check.ipynb).
2. Only correctly classified images were eligible. Every class contributes the same number of images (1,242), split 70/15/15 inside each class.
3. Each image received one attack: a target class that is never its own class, an attack type, and that attack's hyperparameters. The assignment is random but constrained, so target classes, attack types, hyperparameter settings, source classes, and splits are all evenly represented (balance table below).
4. Attacks minimize float32 cross-entropy toward the target over the 10 Imagenette logits. Results are rounded to 8 bits, stored as lossless PNG, decoded, and re-classified.
5. Every attempt is kept, including failed ones. target_hit and label_flipped record the outcome after the PNG round trip. Keeping failures preserves the balance; filter on these flags for success-only subsets (that subset is no longer balanced).

Balance

spread is the largest max − min count difference inside any group.

property	mean_count	spread	allowed
source class — overall	1242.0	0	0
source class — per split	414.0	0	0
target class — overall	1242.0	0	0
target class — per split	414.0	0	0
source → target pair (own class excluded)	138.0	0	1
attack — overall	6210.0	0	6
attack — per split	2070.0	2	8
attack — per source class	621.0	2	8
attack — per target class	621.0	2	8
attack — per split × source	207.0	3	10
attack — per split × target	207.0	4	10
attack — per source × target	69.0	2	10
attack — per split × source × target	23.0	4	10
bim setting — overall	2070.0	0	6
bim setting — per split	690.0	2	8
bim setting — per source class	207.0	4	8
bim setting — per target class	207.0	5	8
bim setting — per source × target	23.0	4	10
pgd setting — overall	690.0	3	6
pgd setting — per split	230.0	4	8
pgd setting — per source class	69.0	3	8
pgd setting — per target class	69.0	4	8
pgd setting — per source × target	7.7	4	10
ε (both attacks) — overall	4140.0	3	6
PGD steps — overall	2070.0	2	6
ε (both attacks) — per split	1380.0	4	8
PGD steps — per split	690.0	3	8
ε (both attacks) — per source class	414.0	4	8
PGD steps — per source class	207.0	3	8
ε (both attacks) — per target class	414.0	3	8
PGD steps — per target class	207.0	3	8

Attack success by setting

config_id	epsilon	pairs	target_hit_rate	label_flip_rate	imagenet_target_hit_rate
bim_eps2	2/255	2070	65.9%	65.9%	38.4%
bim_eps4	4/255	2070	87.2%	87.2%	68.5%
bim_eps8	8/255	2070	94.9%	95.1%	85.4%
pgd_eps2_steps10	2/255	691	74.8%	75.0%	49.8%
pgd_eps2_steps20	2/255	690	95.1%	95.1%	82.9%
pgd_eps2_steps40	2/255	688	99.7%	99.7%	96.1%
pgd_eps4_steps10	4/255	689	92.3%	92.3%	72.7%
pgd_eps4_steps20	4/255	690	98.6%	98.6%	93.2%
pgd_eps4_steps40	4/255	690	100.0%	100.0%	99.7%
pgd_eps8_steps10	8/255	691	96.2%	96.2%	82.8%
pgd_eps8_steps20	8/255	690	99.4%	99.4%	95.7%
pgd_eps8_steps40	8/255	691	100.0%	100.0%	99.4%

Misclassified-class distribution

Predicted class of attacked images whose label flipped. Targets are balanced, so these counts should be roughly even across classes:

predicted class	bim	pgd
tench	507	583
English springer	512	588
cassette player	498	587
chain saw	530	589
church	507	593
French horn	487	580
garbage truck	517	593
gas pump	508	589
golf ball	546	605
parachute	526	601

attack	misclassified_images	min_per_class	max_per_class	largest_class_share	on_assigned_target	uniform_p_value
bim	5138	487	546	10.6%	99.9%	0.837
pgd	5908	580	605	10.2%	100.0%	1.000

ImageNet training-set overlap

About 96% of Imagenette images come from the ImageNet-1k training split, which the classifier was trained on. imagenet_origin tells the two apart (val = never seen in training):

imagenet_origin	pairs	target_hit_rate	label_flip_rate
train	11962	88.8%	88.9%
val	458	90.2%	90.2%

Fields

• Identity: pair_id, split, source_dataset, source_id (path inside the Imagenette archive), imagenette_split, imagenet_origin, true_wnid, true_label, true_class_name
• Images: clean_image, adversarial_image (224×224 RGB PNG)
• Attack: attack, config_id, epsilon_level (ε × 255), epsilon, steps, step_size, random_start, attack_seed, target_label, target_class_name
• Clean result: clean_prediction, clean_prediction_name, clean_confidence, clean_correct, clean_imagenet_top1
• Adversarial result: adversarial_prediction, adversarial_prediction_name, adversarial_confidence, adversarial_target_probability, adversarial_true_probability, adversarial_imagenet_top1
• Outcome: target_hit, label_flipped, target_hit_before_png, label_flipped_before_png, imagenet_target_hit
• Perturbation (8-bit): perturbation_linf_levels, perturbation_linf, perturbation_l2
• Provenance: clean_sha256, adversarial_sha256, model_id, model_revision, device_name

Probabilities and predictions are over the 10 Imagenette classes unless the name contains imagenet. metadata/ holds the frozen attack plan, manifest, generation report, audit report, and figures.

Loading

from datasets import load_dataset

pairs = load_dataset("nimaeb/xai-attack-detection-imagenette", split="train")
hits = pairs.filter(lambda row: row["target_hit"])

Limitations

• One classifier, one source dataset, and two closely related white-box L∞ attacks (BIM is PGD without the random start).
• Targets are restricted to the 10 Imagenette classes. The attack does not steer the 1000-class output.
• All images are in the 8-bit PNG domain.
• Clean accuracy is inflated by the ImageNet training overlap described above.

License and access

Private research artifact with no reuse license. Imagenette is a subset of ImageNet, so ImageNet's terms of access apply to the images. Review licensing before any public release.