<i>CAGFuzz:</i> Coverage-Guided Adversarial Generative Fuzzing Testing for Image-Based Deep Learning Systems

Zhang, Pengcheng; Ren, Bin; Dong, Hai; Dai, Qiyin

doi:10.1109/tse.2021.3124006

Cited by 16 publications

(4 citation statements)

References 37 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For performance evaluation and comparison, we use six different indicators, including accuracy, precision, recall, F1 Score and ROC curve [3], [11], [19], [24], [35], [36]. Among them, accuracy is the proportion of the number of correctly classified sample in the total number of samples.…”

Section: Resultsmentioning

confidence: 99%

Multi-Class Prediction of Mineral Resources Based on Deep Learning

et al. 2022

View full text Add to dashboard Cite

Big data-driven technologies, especially machine learning and deep learning technologies, have been extensively employed in mineral prospectivity prediction. Several approaches have been proposed to learn the deep characteristics of geoscience data, enhance the accuracy of prediction and reduce uncertainty. Nevertheless, the approaches always contain the following two limitations. Firstly, the formation of mineral resources often involves the coupling of multiple factors on a certain spatio-temporal scale, resulting in rare labelled deposits and insufficient number of training samples. Secondly, training Deep Neural Network (DNN) is very challenging. Many approaches are subject to weak interpretability and lack of organic combination with geoscience knowledge. To address these two problems, we propose Geo-Rnet and GCAE (Geological Convolutional Autoencoder). Geo-Rnet is a multi-class mineral prospectivity prediction approach based on improved DNN. GCAE is able to effectively augment multi-disciplinary geoscience data by constructing upon an optimized Convolutional Autoencoder. The experimental results show that most of prospective areas predicted by Geo-Rnet overlap with the labelled mineralization locations, with an average accuracy of 91.1%. In addition, 89.98% of the ore deposits are located in the predicted areas. The results indicate the effectiveness of Geo-Rnet and GCAE for multi-class prediction of mineral resources. Finally, we classify the target area into several mineral prospectiviy areas according to their different mineral types. The research provides an innovative approach for mineral prospectivity prediction in the target area.INDEX TERMS Deep neural network, Geo-Rnet, multi-class, mineral prediction.

show abstract

Section: Resultsmentioning

confidence: 99%

Multi-Class Prediction of Mineral Resources Based on Deep Learning

et al. 2022

View full text Add to dashboard Cite

show abstract

“…As discussed in our Introduction, white-box testing [12], [16], [20], [25], [35], [37] has been extensively researched, but needs internal model access, which is not always realistic in practice. Therefore, we will adopt a black-box testing approach instead, which purely focuses on modifying a system's input (in our case, an image) to trigger undesired changes in the system's output (in our case, the object classification for the input image).…”

Section: Related Workmentioning

confidence: 99%

“…Existing approaches to adversarial example generation can be classified into white-box and black-box methods. White-box approaches [12], [16], [20], [25], [35], [37] require access to the model under test (i.e. the model architecture, neuron weight values, and gradients).…”

Section: Introductionmentioning

confidence: 99%

On the Strengths of Pure Evolutionary Algorithms in Generating Adversarial Examples

Bartlett,

Liem,

Panichella

2023

2023 IEEE/ACM International Workshop on Search-Based and Fuzz Testing (SBFT)

View full text Add to dashboard Cite

Deep learning (DL) models are known to be highly accurate, yet vulnerable to adversarial examples. While earlier research focused on generating adversarial examples using whitebox strategies, later research focused on black-box strategies, as models often are not accessible to external attackers. Prior studies showed that black-box approaches based on approximate gradient descent algorithms combined with meta-heuristic search (i.e., the BMI-FGSM algorithm) outperform previously proposed white-and black-box strategies. In this paper, we propose a novel black-box approach purely based on differential evolution (DE), i.e., without using any gradient approximation method. In particular, we propose two variants of a customized DE with customized variation operators: (1) a single-objective (Pixel-SOO) variant generating attacks that fool DL models, and (2) a multi-objective variant (Pixel-MOO) that also minimizes the number of changes in generated attacks. Our preliminary study on five canonical image classification models shows that Pixel-SOO and Pixel-MOO are more effective than the state-of-the-art BMI-FGSM in generating adversarial attacks. Furthermore, Pixel-SOO is faster than Pixel-MOO, while the latter produces subtler attacks than its single-objective variant.

show abstract

“…One simple and commonly used strategy is random selection, but it does not use any information of testing process or DL systems. To enhance the random strategy, the idea of "recency-aware" was used to select the seed that induces the new coverage [9,7,4,10,11,12], which corresponds to the observation that if a seed covers a branch, the following branches are more likely to be covered due to the hierarchical relationship between branches. In addition, the idea of "frequencyaware" was used to probabilistically select a seed based on the number of times it has been mutated: if a seed has already been picked many times, it has a lower probability of being selected again [4,10].…”

Section: Related Workmentioning

confidence: 99%

DeepController: Feedback-Directed Fuzzing for Deep Learning Systems

Dai¹,

Sun²,

Liu³

2022

Proceedings of the 34th International Conference on Software Engineering and Knowledge Engineering

View full text Add to dashboard Cite

Deep learning (DL) systems are increasingly adopted in various fields, while fatal failures are still inevitable in them. One mainstream testing approach for DL is fuzzing, which can generate a large amount of semi-random yet syntactically valid test cases. Previous studies on fuzzing are mainly focused on selecting "quality" seeds or using "good" mutation strategies. In this paper, we attempt to improve the performance of fuzzing from a different perspective. A new fuzzer, namely DeepController, is accordingly developed, which makes use of the feedback information obtained in the test execution process to dynamically select seeds and mutation strategies. DeepController is evaluated through empirical studies on three datasets and eight DL models. The experimental results show that, with the same number of seeds, DeepController can generate more adversarial inputs and achieve higher neuron coverage than the state-of-the-art testing techniques for DL systems.

show abstract

CAGFuzz: Coverage-Guided Adversarial Generative Fuzzing Testing for Image-Based Deep Learning Systems

Cited by 16 publications

References 37 publications

Multi-Class Prediction of Mineral Resources Based on Deep Learning

Multi-Class Prediction of Mineral Resources Based on Deep Learning

On the Strengths of Pure Evolutionary Algorithms in Generating Adversarial Examples

DeepController: Feedback-Directed Fuzzing for Deep Learning Systems

Contact Info

Product

Resources

About