Deep neural network (DNN) testing is crucial for the reliability and safety of critical systems, where failures can have severe consequences. Although various techniques have been developed to create robustness test suites, requirements-based testing for DNNs remains largely unexplored - yet such tests are recognized as an essential component of software validation of critical systems. In this work, we propose a requirements-based test suite generation method that uses structured natural language requirements formulated in a semantic feature space to create test suites by prompting text-conditional latent diffusion models with the requirement precondition and then using the associated postcondition to define a test oracle to judge outputs of the DNN under test. We investigate the approach using fine-tuned variants of pre-trained generative models. Our experiments on the MNIST, CelebA-HQ, ImageNet, and autonomous car driving datasets demonstrate that the generated test suites are realistic, diverse, consistent with preconditions, and capable of revealing faults.
The generation of synthetic inputs via simulators driven by search algorithms is essential for cost-effective testing of Deep Neural Network (DNN) components for safety-critical systems. However, in many applications, simulators are unable to produce the ground-truth data needed for automated test oracles and to guide the search process. To tackle this issue, we propose an approach for the generation of inputs for computer vision DNNs that integrates a generative network to ensure simulator fidelity and employs heuristic-based search fitnesses that leverage transformation consistency, noise resistance, surprise adequacy, and uncertainty estimation. We compare the performance of our fitnesses with that of a traditional fitness function leveraging ground truth; further, we assess how the integration of a GAN not leveraging the ground truth impacts on test and retraining effectiveness. Our results suggest that leveraging transformation consistency is the best option to generate inputs for both DNN testing and retraining; it maximizes input diversity, spots the inputs leading to worse DNN performance, and leads to best DNN performance after retraining. Besides enabling simulator-based testing in the absence of ground truth, our findings pave the way for testing solutions that replace costly simulators with diffusion and large language models, which might be more affordable than simulators, but cannot generate ground-truth data.
Surprise Adequacy (SA) has been widely studied as a test adequacy metric that can effectively guide software engineers towards inputs that are more likely to reveal unexpected behaviour of Deep Neural Networks (DNNs). Intuitively, SA is an out-of-distribution metric that quantifies the dissimilarity between the given input and the training data: if a new input is very different from those seen during training, the DNN is more likely to behave unexpectedly against the input. While SA has been widely adopted as a test prioritization method, its major weakness is the fact that the computation of the metric requires access to the training dataset, which is often not allowed in real-world use cases. We present DANDI, a technique that generates a surrogate input distribution using Stable Diffusion to compute SA values without requiring the original training data. An empirical evaluation of DANDI applied to image classifiers for CIFAR10 and ImageNet-1K shows that SA values computed against synthetic data are highly correlated with the values computed against the training data, with Spearman Rank correlation value of 0.852 for ImageNet-1K and 0.881 for CIFAR-10. Further, we show that SA value computed by DANDI achieves can prioritize inputs as effectively as those computed using the training data, when testing DNN models mutated by DeepMutation. We believe that DANDI can significantly improve the usability of SA for practical DNN testing.
Deep Neural Networks (DNNs) for Autonomous Driving Systems (ADS) are typically trained on real-world images and tested using synthetic simulator images. This approach results in training and test datasets with dissimilar distributions, which can potentially lead to erroneously decreased test accuracy. To address this issue, the literature suggests applying domain-to-domain translators to test datasets to bring them closer to the training datasets. However, translating images used for testing may unpredictably affect the reliability, effectiveness and efficiency of the testing process. Hence, this paper investigates the following questions in the context of ADS: Could translators reduce the effectiveness of images used for ADS-DNN testing and their ability to reveal faults in ADS-DNNs? Can translators result in excessive time overhead during simulation-based testing? To address these questions, we consider three domain-to-domain translators: CycleGAN and neural style transfer, from the literature, and SAEVAE, our proposed translator. Our results for two critical ADS tasks -- lane keeping and object detection -- indicate that translators significantly narrow the gap in ADS test accuracy caused by distribution dissimilarities between training and test data, with SAEVAE outperforming the other two translators. We show that, based on the recent diversity, coverage, and fault-revealing ability metrics for testing deep-learning systems, translators do not compromise the diversity and the coverage of test data, nor do they lead to revealing fewer faults in ADS-DNNs. Further, among the translators considered, SAEVAE incurs a negligible overhead in simulation time and can be efficiently integrated into simulation-based testing. Finally, we show that translators increase the correlation between offline and simulation-based testing results, which can help reduce the cost of simulation-based testing.
Evaluating the behavioral boundaries of deep learning (DL) systems is crucial for understanding their reliability across diverse, unseen inputs. Existing solutions fall short as they rely on untargeted random, model- or latent-based perturbations, due to difficulties in generating controlled input variations. In this work, we introduce Mimicry, a novel black-box test generator for fine-grained, targeted exploration of DL system boundaries. Mimicry performs boundary testing by leveraging the probabilistic nature of DL outputs to identify promising directions for exploration. It uses style-based GANs to disentangle input representations into content and style components, enabling controlled feature mixing to approximate the decision boundary. We evaluated Mimicry's effectiveness in generating boundary inputs for five widely used DL image classification systems of increasing complexity, comparing it to two baseline approaches. Our results show that Mimicry consistently identifies inputs closer to the decision boundary. It generates semantically meaningful boundary test cases that reveal new functional (mis)behaviors, while the baselines produce mainly corrupted or invalid inputs. Thanks to its enhanced control over latent space manipulations, Mimicry remains effective as dataset complexity increases, maintaining competitive diversity and higher validity rates, confirmed by human assessors.
Deep Neural Networks (DNN) are core components for classification and regression tasks of many software systems. Companies incur in high costs for testing DNN with datasets representative of the inputs expected in operation, as these need to be manually labelled. The challenge is to select a representative set of test inputs as small as possible to reduce the labelling cost, while sufficing to yield unbiased high-confidence estimates of the expected DNN accuracy. At the same time, testers are interested in exposing as many DNN mispredictions as possible to improve the DNN, ending up in the need for techniques pursuing a threefold aim: small dataset size, trustworthy estimates, mispredictions exposure. This study presents DeepSample, a family of DNN testing techniques for cost-effective accuracy assessment based on probabilistic sampling. We investigate whether, to what extent, and under which conditions probabilistic sampling can help to tackle the outlined challenge. We implement five new sampling-based testing techniques, and perform a comprehensive comparison of such techniques and of three further state-of-the-art techniques for both DNN classification and regression tasks. Results serve as guidance for best use of sampling-based testing for faithful and high-confidence estimates of DNN accuracy in operation at low cost.
Despite their unprecedented success, DNNs are notoriously fragile to small shifts in data distribution, demanding effective testing techniques that can assess their dependability. Despite recent advances in DNN testing, there is a lack of systematic testing approaches that assess the DNN's capability to generalise and operate comparably beyond data in their training distribution. We address this gap with DeepKnowledge, a systematic testing methodology for DNN-based systems founded on the theory of knowledge generalisation, which aims to enhance DNN robustness and reduce the residual risk of 'black box' models. Conforming to this theory, DeepKnowledge posits that core computational DNN units, termed Transfer Knowledge neurons, can generalise under domain shift. DeepKnowledge provides an objective confidence measurement on testing activities of DNN given data distribution shifts and uses this information to instrument a generalisation-informed test adequacy criterion to check the transfer knowledge capacity of a test set. Our empirical evaluation of several DNNs, across multiple datasets and state-of-the-art adversarial generation techniques demonstrates the usefulness and effectiveness of DeepKnowledge and its ability to support the engineering of more dependable DNNs. We report improvements of up to 10 percentage points over state-of-the-art coverage criteria for detecting adversarial attacks on several benchmarks, including MNIST, SVHN, and CIFAR.
Motivated by the success of traditional software testing, numerous diversity measures have been proposed for testing deep neural networks (DNNs). In this study, we propose a shift in perspective, advocating for the consideration of DNN testing as directed testing problems rather than diversity-based testing tasks. We note that the objective of testing DNNs is specific and well-defined: identifying inputs that lead to misclassifications. Consequently, a more precise testing approach is to prioritize inputs with a higher potential to induce misclassifications, as opposed to emphasizing inputs that enhance "diversity." We derive six directed metrics for DNN testing. Furthermore, we conduct a careful analysis of the appropriate scope for each metric, as applying metrics beyond their intended scope could significantly diminish their effectiveness. Our evaluation demonstrates that (1) diversity metrics are particularly weak indicators for identifying buggy inputs resulting from small input perturbations, and (2) our directed metrics consistently outperform diversity metrics in revealing erroneous behaviors of DNNs across all scenarios.
In this paper, we propose a new Deep Neural Network (DNN) testing algorithm called the Constrained Gradient Descent (CGD) method, and an implementation we call CGDTest aimed at exposing security and robustness issues such as adversarial robustness and bias in DNNs. Our CGD algorithm is a gradient-descent (GD) method, with the twist that the user can also specify logical properties that characterize the kinds of inputs that the user may want. This functionality sets CGDTest apart from other similar DNN testing tools since it allows users to specify logical constraints to test DNNs not only for $\ell_p$ ball-based adversarial robustness but, more importantly, includes richer properties such as disguised and flow adversarial constraints, as well as adversarial robustness in the NLP domain. We showcase the utility and power of CGDTest via extensive experimentation in the context of vision and NLP domains, comparing against 32 state-of-the-art methods over these diverse domains. Our results indicate that CGDTest outperforms state-of-the-art testing tools for $\ell_p$ ball-based adversarial robustness, and is significantly superior in testing for other adversarial robustness, with improvements in PAR2 scores of over 1500% in some cases over the next best tool. Our evaluation shows that our CGD method outperforms competing methods we compared against in terms of expressibility (i.e., a rich constraint language and concomitant tool support to express a wide variety of properties), scalability (i.e., can be applied to very large real-world models with up to 138 million parameters), and generality (i.e., can be used to test a plethora of model architectures).
Deep neural network (DNN) models, including those used in safety-critical domains, need to be thoroughly tested to ensure that they can reliably perform well in different scenarios. In this article, we provide an overview of structural coverage metrics for testing DNN models, including neuron coverage (NC), k-multisection neuron coverage (kMNC), top-k neuron coverage (TKNC), neuron boundary coverage (NBC), strong neuron activation coverage (SNAC) and modified condition/decision coverage (MC/DC). We evaluate the metrics on realistic DNN models used for perception tasks (including LeNet-1, LeNet-4, LeNet-5, and ResNet20) as well as on networks used in autonomy (TaxiNet). We also provide a tool, DNNCov, which can measure the testing coverage for all these metrics. DNNCov outputs an informative coverage report to enable researchers and practitioners to assess the adequacy of DNN testing, compare different coverage measures, and to more conveniently inspect the model's internals during testing.
Deep learning has recently been widely applied to many applications across different domains, e.g., image classification and audio recognition. However, the quality of Deep Neural Networks (DNNs) still raises concerns in the practical operational environment, which calls for systematic testing, especially in safety-critical scenarios. Inspired by software testing, a number of structural coverage criteria are designed and proposed to measure the test adequacy of DNNs. However, due to the blackbox nature of DNN, the existing structural coverage criteria are difficult to interpret, making it hard to understand the underlying principles of these criteria. The relationship between the structural coverage and the decision logic of DNNs is unknown. Moreover, recent studies have further revealed the non-existence of correlation between the structural coverage and DNN defect detection, which further posts concerns on what a suitable DNN testing criterion should be. In this paper, we propose the interpretable coverage criteria through constructing the decision structure of a DNN. Mirroring the control flow graph of the traditional program, we first extract a decision graph from a DNN based on its interpretation, where a path of the decision graph represents a decision logic of the DNN. Based on the control flow and data flow of the decision graph, we propose two variants of path coverage to measure the adequacy of the test cases in exercising the decision logic. The higher the path coverage, the more diverse decision logic the DNN is expected to be explored. Our large-scale evaluation results demonstrate that: the path in the decision graph is effective in characterizing the decision of the DNN, and the proposed coverage criteria are also sensitive with errors including natural errors and adversarial examples, and strongly correlated with the output impartiality.
Deep Neural Networks (DNNs) have been extensively used in many areas including image processing, medical diagnostics, and autonomous driving. However, DNNs can exhibit erroneous behaviours that may lead to critical errors, especially when used in safety-critical systems. Inspired by testing techniques for traditional software systems, researchers have proposed neuron coverage criteria, as an analogy to source code coverage, to guide the testing of DNN models. Despite very active research on DNN coverage, several recent studies have questioned the usefulness of such criteria in guiding DNN testing. Further, from a practical standpoint, these criteria are white-box as they require access to the internals or training data of DNN models, which is in many contexts not feasible or convenient. In this paper, we investigate black-box input diversity metrics as an alternative to white-box coverage criteria. To this end, we first select and adapt three diversity metrics and study, in a controlled manner, their capacity to measure actual diversity in input sets. We then analyse their statistical association with fault detection using four datasets and five DNN models. We further compare diversity with state-of-the-art white-box coverage criteria. Our experiments show that relying on the diversity of image features embedded in test input sets is a more reliable indicator than coverage criteria to effectively guide the testing of DNNs. Indeed, we found that one of our selected black-box diversity metrics far outperforms existing coverage criteria in terms of fault-revealing capability and computational time. Results also confirm the suspicions that state-of-the-art coverage metrics are not adequate to guide the construction of test input sets to detect as many faults as possible with natural inputs.
The success of deep neural networks (DNNs) in real-world applications has benefited from abundant pre-trained models. However, the backdoored pre-trained models can pose a significant trojan threat to the deployment of downstream DNNs. Numerous backdoor detection methods have been proposed but are limited to two aspects: (1) high sensitivity on trigger size, especially on stealthy attacks (i.e., blending attacks and defense adaptive attacks); (2) rely heavily on benign examples for reverse engineering. To address these challenges, we empirically observed that trojaned behaviors triggered by various trojan attacks can be attributed to the trojan path, composed of top-$k$ critical neurons with more significant contributions to model prediction changes. Motivated by it, we propose CatchBackdoor, a detection method against trojan attacks. Based on the close connection between trojaned behaviors and trojan path to trigger errors, CatchBackdoor starts from the benign path and gradually approximates the trojan path through differential fuzzing. We then reverse triggers from the trojan path, to trigger errors caused by diverse trojaned attacks. Extensive experiments on MINST, CIFAR-10, and a-ImageNet datasets and 7 models (LeNet, ResNet, and VGG) demonstrate the superiority of CatchBackdoor over the state-of-the-art methods, in terms of (1) \emph{effective} - it shows better detection performance, especially on stealthy attacks ($\sim$ $\times$ 2 on average); (2) \emph{extensible} - it is robust to trigger size and can conduct detection without benign examples.
Deep neural networks (DNNs) often accept high-dimensional media data (e.g., photos, text, and audio) and understand their perceptual content (e.g., a cat). To test DNNs, diverse inputs are needed to trigger mis-predictions. Some preliminary works use byte-level mutations or domain-specific filters (e.g., foggy), whose enabled mutations may be limited and likely error-prone. SOTA works employ deep generative models to generate (infinite) inputs. Also, to keep the mutated inputs perceptually valid (e.g., a cat remains a "cat" after mutation), existing efforts rely on imprecise and less generalizable heuristics. This study revisits two key objectives in media input mutation - perception diversity (DIV) and validity (VAL) - in a rigorous manner based on manifold, a well-developed theory capturing perceptions of high-dimensional media data in a low-dimensional space. We show important results that DIV and VAL inextricably bound each other, and prove that SOTA generative model-based methods fundamentally fail to mutate real-world media data (either sacrificing DIV or VAL). In contrast, we discuss the feasibility of mutating real-world media data with provably high DIV and VAL based on manifold. We concretize the technical solution of mutating media data of various formats (images, audios, text) via a unified manner based on manifold. Specifically, when media data are projected into a low-dimensional manifold, the data can be mutated by walking on the manifold with certain directions and step sizes. When contrasted with the input data, the mutated data exhibit encouraging DIV in the perceptual traits (e.g., lying vs. standing dog) while retaining reasonably high VAL (i.e., a dog remains a dog). We implement our techniques in DEEPWALK for testing DNNs. DEEPWALK outperforms prior methods in testing comprehensiveness and can find more error-triggering inputs with higher quality.
Various deep neural network (DNN) coverage criteria have been proposed to assess DNN test inputs and steer input mutations. The coverage is characterized via neurons having certain outputs, or the discrepancy between neuron outputs. Nevertheless, recent research indicates that neuron coverage criteria show little correlation with test suite quality. In general, DNNs approximate distributions, by incorporating hierarchical layers, to make predictions for inputs. Thus, we champion to deduce DNN behaviors based on its approximated distributions from a layer perspective. A test suite should be assessed using its induced layer output distributions. Accordingly, to fully examine DNN behaviors, input mutation should be directed toward diversifying the approximated distributions. This paper summarizes eight design requirements for DNN coverage criteria, taking into account distribution properties and practical concerns. We then propose a new criterion, NeuraL Coverage (NLC), that satisfies all design requirements. NLC treats a single DNN layer as the basic computational unit (rather than a single neuron) and captures four critical properties of neuron output distributions. Thus, NLC accurately describes how DNNs comprehend inputs via approximated distributions. We demonstrate that NLC is significantly correlated with the diversity of a test suite across a number of tasks (classification and generation) and data formats (image and text). Its capacity to discover DNN prediction errors is promising. Test input mutation guided by NLC results in a greater quality and diversity of exposed erroneous behaviors.
Existing methods for testing DNNs solve the oracle problem by constraining the raw features (e.g. image pixel values) to be within a small distance of a dataset example for which the desired DNN output is known. But this limits the kinds of faults these approaches are able to detect. In this paper, we introduce a novel DNN testing method that is able to find faults in DNNs that other methods cannot. The crux is that, by leveraging generative machine learning, we can generate fresh test inputs that vary in their high-level features (for images, these include object shape, location, texture, and colour). We demonstrate that our approach is capable of detecting deliberately injected faults as well as new faults in state-of-the-art DNNs, and that in both cases, existing methods are unable to find these faults.
This paper focuses on the domain generalization task where domain knowledge is unavailable, and even worse, only samples from a single domain can be utilized during training. Our motivation originates from the recent progresses in deep neural network (DNN) testing, which has shown that maximizing neuron coverage of DNN can help to explore possible defects of DNN (i.e., misclassification). More specifically, by treating the DNN as a program and each neuron as a functional point of the code, during the network training we aim to improve the generalization capability by maximizing the neuron coverage of DNN with the gradient similarity regularization between the original and augmented samples. As such, the decision behavior of the DNN is optimized, avoiding the arbitrary neurons that are deleterious for the unseen samples, and leading to the trained DNN that can be better generalized to out-of-distribution samples. Extensive studies on various domain generalization tasks based on both single and multiple domain(s) setting demonstrate the effectiveness of our proposed approach compared with state-of-the-art baseline methods. We also analyze our method by conducting visualization based on network dissection. The results further provide useful evidence on the rationality and effectiveness of our approach.
The reliability of software that has a Deep Neural Network (DNN) as a component is urgently important today given the increasing number of critical applications being deployed with DNNs. The need for reliability raises a need for rigorous testing of the safety and trustworthiness of these systems. In the last few years, there have been a number of research efforts focused on testing DNNs. However the test generation techniques proposed so far lack a check to determine whether the test inputs they are generating are valid, and thus invalid inputs are produced. To illustrate this situation, we explored three recent DNN testing techniques. Using deep generative model based input validation, we show that all the three techniques generate significant number of invalid test inputs. We further analyzed the test coverage achieved by the test inputs generated by the DNN testing techniques and showed how invalid test inputs can falsely inflate test coverage metrics. To overcome the inclusion of invalid inputs in testing, we propose a technique to incorporate the valid input space of the DNN model under test in the test generation process. Our technique uses a deep generative model-based algorithm to generate only valid inputs. Results of our empirical studies show that our technique is effective in eliminating invalid tests and boosting the number of valid test inputs generated.
The widespread application of deep neural network (DNN) techniques is being challenged by adversarial examples, the legitimate input added with imperceptible and well-designed perturbations that can fool DNNs easily in the DNN testing/deploying stage. Previous adversarial example generation algorithms for adversarial white-box attacks used Jacobian gradient information to add perturbations. This information is too imprecise and inexplicit, which will cause unnecessary perturbations when generating adversarial examples. This paper aims to address this issue. We first propose to apply a more informative and distilled gradient information, namely integrated gradient, to generate adversarial examples. To further make the perturbations more imperceptible, we propose to employ the restriction combination of $L_0$ and $L_1/L_2$ secondly, which can restrict the total perturbations and perturbation points simultaneously. Meanwhile, to address the non-differentiable problem of $L_1$, we explore a proximal operation of $L_1$ thirdly. Based on these three works, we propose two Integrated gradient based White-box Adversarial example generation algorithms (IWA): IFPA and IUA. IFPA is suitable for situations where there are a determined number of points to be perturbed. IUA is suitable for situations where no perturbation point number is preset in order to obtain more adversarial examples. We verify the effectiveness of the proposed algorithms on both structured and unstructured datasets, and we compare them with five baseline generation algorithms. The results show that our proposed algorithms do craft adversarial examples with more imperceptible perturbations and satisfactory crafting rate. $L_2$ restriction is more suitable for unstructured dataset and $L_1$ restriction performs better in structured dataset.
To analyse local robustness properties of deep neural networks (DNNs), we present a practical framework from a model learning perspective. Based on black-box model learning with scenario optimisation, we abstract the local behaviour of a DNN via an affine model with the probably approximately correct (PAC) guarantee. From the learned model, we can infer the corresponding PAC-model robustness property. The innovation of our work is the integration of model learning into PAC robustness analysis: that is, we construct a PAC guarantee on the model level instead of sample distribution, which induces a more faithful and accurate robustness evaluation. This is in contrast to existing statistical methods without model learning. We implement our method in a prototypical tool named DeepPAC. As a black-box method, DeepPAC is scalable and efficient, especially when DNNs have complex structures or high-dimensional inputs. We extensively evaluate DeepPAC, with 4 baselines (using formal verification, statistical methods, testing and adversarial attack) and 20 DNN models across 3 datasets, including MNIST, CIFAR-10, and ImageNet. It is shown that DeepPAC outperforms the state-of-the-art statistical method PROVERO, and it achieves more practical robustness analysis than the formal verification tool ERAN. Also, its results are consistent with existing DNN testing work like DeepGini.