3 results
Now showing1 - 3 of 3
- Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , Efficient adversarial generation through selective SBFL guidance and automated sub-model strategyFast Gradient Sign Method (FGSM) is an effective method for generating adversarial inputs for neural networks, but it is memory-intensive. DeepFault reduces the memory costs of FGSM by transferring Spectrum-Based Fault Localization (SBFL) to neural networks. SBFL is a technique traditionally using the execution trace of a program to iden-tify suspicious code locations that are likely to contain faults. DeepFault employs SBFL to identify neurons in a neural network that are likely to be responsible for misclassifications to guide FGSM. We propose an adjustable SBFL-guided adversarial input generation approach that addresses both memory and runtime efficiency challenges for sequential classification models. Our contributions include: (1) an automated and adjustable sub-model strategy that selects smaller model configurations for input synthesis, reducing both memory overhead and computational effort, (2) vector-based hit-spectrum computation using GPU-accelerated operations to significantly reduce computation time (3) targeted adversarial input generation enabling modifications in favor of specific target classes, and (4) comprehensive experimental evaluation on contemporary sequential neural network architectures. Our evaluation demonstrates substantial improvements over existing methods: up to 4.7× memory reduction compared to FGSM and up to 301× speedup over DeepFault while maintaining comparable adversarial input quality. All experiments complete within tens of seconds, demonstrating practical scalability for contemporary models.Publicationtype: Conference Paper11 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , Efficient Hit-Spectrum-Guided Fast Gradient Sign Method: An Adjustable Approach with Memory and Runtime OptimizationsFast Gradient Sign Method (FGSM) is an effective method for generating adversarial inputs for neural networks, but it is memory-intensive. DeepFault reduces the memory costs of FGSM by transferring Spectrum-Based Fault Localization to neural networks. SBFL is a technique traditionally using the execution trace of a program to identify suspicious code locations that are likely to contain faults. DeepFault employs SBFL to identify neurons in a neural network that are likely to be responsible for misclassifications to guide FGSM. We propose an adjustable hit-spectrum-guided FGSM approach applying a sub-model strategy to avoid gradient ascent evaluation over the entire model. Additionally, we alter DeepFault’s hit-spectrum computation to be vector-based to allow parallelization of computation, and we modify the hit spectrum to depend on a specific class to allow targeted adversarial input generation. We conduct an experimental evaluation on image classification models showing how our app roach allows trading off effectiveness of adversarial input generation with reduced runtimes while maintaining scalability regarding larger models, with maximum runtimes on the order of tens of seconds. For larger sample sizes, our approach reduces runtimes to fractions of 1/300 and less compared to DeepFault. When processing larger models, it requires only one-third of FGSM’s memory usage.Publicationtype: Conference Paper20 - Some of the metrics are blocked by yourconsent settings
Item type:Publication, Access status: without files , Repairing neural networks for image classification problems using spectrum-based fault localizationFault localization in neural networks develops methods to improve a model's accuracy, resilience, and performance. Improving a model's accuracy without extensive gradient descent training loops has not been a goal, yet. In this paper, we propose a novel model repair approach identifying suspicious neurons to be selected for modifications of greater impact upon a single iteration avoiding extensive training processes with the gradient descent method yielding potential overfitting. Our approach uses Spectrum-Based Fault Localization to guide the repair process. Spectrum-Based Fault Localization is an established method that assigns software code a suspiciousness value indicating a likelihood for fault presence depending on test suite outcomes and the associated code's execution. We employ hit-spectrum analysis to assign neurons a suspiciousness score. Our algorithm selects especially suspicious neurons for repair by reducing their activation values, while increasing the activation values of less suspicious neurons. Experiments conducted on convolutional layer and dense layer image classification models confirm that, overall, our approach leads to better mean accuracies. In 2 scenarios, the repair approach has led to small accuracy improvements by increasing the corresponding models' mean accuracies by 0.68 and 0.46 percentage points respectively. Remarkably, our repair approach has also led to significant accuracy improvements by increasing a specific model's accuracy by 15.73 percentage points. In conclusion, our method provides a promising alternative to additional training epochs running the risk of overfitting or depending on larger databases for the training process.Publicationtype: Conference Paper20