A Novel Semi-Supervised Adversarially Learned Meta-Classifier for Detecting Neural Trojan Attacks
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) are highly vulnerable to neural Trojan attacks. To carry out such an attack, an adversary retrains a DNN with poisoned data or modifies its parameters to produce incorrect output. These attacks can remain unnoticed until triggered by a specific pattern in the input, making detection challenging. In this article, we propose a novel semi-supervised adversarially learned meta-classifier (SESALME) to detect if a target model has been trojaned. Unlike previous Trojan detection methods, SESALME assumes that the defender has no knowledge of the attack mechanisms, and no access to training data, poisoned data, or parameters/layers of a target model. In the absence of poisoned data and knowledge of the attack mechanisms, we use a set of shadow models to emulate normal behavior of the target model. Having learned the normal behavior of the target model, SESALME then uses one-class learning, implemented within a semi-supervised generative adversarial network (GAN), to detect abnormal behavior of a model to be investigated, if any. Behavior that deviates from the learned normal behavior indicates a high likelihood that the model is trojaned. We compare the performance of SESALME with that of state-of-the-art neural Trojan detectors using popular datasets such as MNIST, CIFAR-10, and SC. Experimental results show that SESALME outperforms state-of-the-art Trojan detection methods in terms of detection performance and inference time in almost all cases, while being attack-agnostic and requiring no access to training data, poisoned data, or parameters of the target model.
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 | 0.001 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.000 | 0.002 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.003 | 0.001 |
| 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