Gradual poisoning of a chest x-ray convolutional neural network with an adversarial attack and AI explainability methods
Bibliographic record
Abstract
Given artificial intelligence's transformative effects, studying safety is important to ensure it is implemented in a beneficial way. Convolutional neural networks are used in radiology research for prediction but can be corrupted through adversarial attacks. This study investigates the effect of an adversarial attack, through poisoned data. To improve generalizability, we create a generic ResNet pneumonia classification model and then use it as an example by subjecting it to BadNets adversarial attacks. The study uses various poisoned datasets of different compositions (2%, 16.7% and 100% ratios of poisoned data) and two different test sets (a normal set of test data and one that contained poisoned images) to study the effects of BadNets. To provide a visual effect of the progressing corruption of the models, SHapley Additive exPlanations (SHAP) were used. As corruption progressed, interval analysis revealed that performance on a valid test set decreased while the model learned to predict better on a poisoned test set. SHAP visualization showed focus on the trigger. In the 16.7% poisoned model, SHAP focus did not fixate on the trigger in the normal test set. Minimal effects were seen in the 2% model. SHAP visualization showed decreasing performance was correlated with increasing focus on the trigger. Corruption could potentially be masked in the 16.7% model unless subjected specifically to poisoned data. A minimum threshold for corruption may exist. The study demonstrates insights that can be further studied in future work and with future models. It also identifies areas of potential intervention for safeguarding models against adversarial attacks.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.004 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".