Inconspicuous Data Augmentation Based Backdoor Attack on Deep Neural Networks
Bibliographic record
Abstract
With new applications made possible by the fusion of edge computing and artificial intelligence (AI) technologies, the global market capitalization of edge AI has risen tremendously in recent years. Deployment of pre-trained deep neural network (DNN) models on edge computing platforms, however, does not alleviate the fundamental trust assurance issue arising from the lack of interpretability of end-to-end DNN solutions. The most notorious threat of DNNs is the backdoor attack. Most backdoor attacks require a relatively large injection rate (≈ 10%) to achieve a high attack success rate. The trigger patterns are not always stealthy and can be easily detected or removed by backdoor detectors. Moreover, these attacks are only tested on DNN models implemented on general-purpose computing platforms. This paper proposes to use data augmentation for backdoor attacks to increase the stealth, attack success rate, and robustness. Different data augmentation techniques are applied independently on three color channels to embed a composite trigger. The data augmentation strength is tuned based on the Gradient Magnitude Similarity Deviation, which is used to objectively assess the visual imperceptibility of the poisoned samples. A rich set of composite triggers can be created for different dirty labels. The proposed attacks are evaluated on pre-activation ResNet18 trained with CIFAR-10 and GTSRB datasets, and EfficientNet-B0 trained with adapted 10-class ImageNet dataset. A high attack success rate of above 97% with only 1% injection rate is achieved on these DNN models implemented on both general-purpose computing platforms and Intel Neural Compute Stick 2 edge AI device. The accuracy loss of the poisoned DNNs on benign inputs is kept below 0.6%. The proposed attack is also tested to be resilient to state-of-the-art backdoor defense methods.
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 machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.003 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.001 | 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 source (direct Gemma or distilled Codex), 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".