Self-Checking Deep Neural Networks for Anomalies and Adversaries in Deployment
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Deep Neural Networks (DNNs) have been widely adopted, yet DNN models are surprisingly unreliable, which raises significant concerns about their use in critical domains. In this work, we propose that runtime DNN mistakes can be quickly detected and properly dealt with <i>in deployment</i>, especially in settings like self-driving vehicles. Just as software engineering (SE) community has developed effective mechanisms and techniques to monitor and check programmed components, our previous work, SelfChecker, is designed to monitor and correct DNN predictions given unintended abnormal test data. SelfChecker triggers an alarm if the decisions given by the internal layer features of the model are inconsistent with the final prediction and provides <i>advice</i> in the form of an alternative prediction. In this paper, we extend SelfChecker to the security domain. Specifically, we describe SelfChecker++, which we designed to target both <i>unintended</i> abnormal test data and <i>intended</i> adversarial samples. Technically, we develop a technique which can transform any runtime inputs triggering alarms into semantically equivalent inputs, then we feed those transformed inputs to the model. Such runtime transformation nullifies any intended crafted samples, making the model immune to adversarial attacks that craft adversarial samples. We evaluated the alarm accuracy of SelfChecker++ on three DNN models and four popular image datasets, and found that SelfChecker++ triggers correct alarms on 63.10% of wrong DNN predictions, and triggers false alarms on 5.77% of correct DNN predictions. We also evaluated the effectiveness of SelfChecker++ in detecting adversarial examples and found it detects on average 70.09% of such samples with advice accuracy that is 20.89% higher than the original DNN model and 18.37% higher than SelfChecker.
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.
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 | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.001 |
| 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 it