No abstract

Deep neural networks (DNNs) have a wide range of applications, and software employing them must be thoroughly tested, especially in safety-critical domains. However, traditional software test coverage metrics cannot be applied directly to DNNs. In this paper, inspired by the MC/DC coverage criterion, we propose a family of four novel test coverage criteria that are tailored to structural features of DNNs and their semantics. We validate the criteria by demonstrating that test inputs that are generated with guidance by our proposed coverage criteria are able to capture undesired behaviours in a DNN. Test cases are generated using a symbolic approach and a gradient-based heuristic search. By comparing them with existing methods, we show that our criteria achieve a balance between their ability to find bugs (proxied using adversarial examples and correlation with functional coverage) and the computational cost of test input generation. Our experiments are conducted on state-of-the-art DNNs obtained using popular open source datasets, including MNIST, CIFAR-10 and ImageNet.

Electrical activation and redistribution of 500 eV boron implants in preamorphized silicon after nonmelt laser annealing at 1150°C and isochronal rapid thermal postannealing are reported. Under the thermal conditions used for a nonmelt laser at 1150°C, a substantial residue of end-of-range defects remained after one laser scan but these were mainly dissolved within ten scans. The authors find dramatic boron deactivation and transient enhanced diffusion after postannealing the one-scan samples, but very little in the five-and ten-scan samples. The results show that end-of-range defect removal during nonmelt laser annealing is an achievable method for the stabilization of highly activated boron profiles in preamorphized silicon. © 2006 American Institute of Physics. ͓DOI: 10.1063/1.2385215͔The continued downscaling of complementary metaloxide-semiconductor devices requires ultrashallow and abrupt source/drain extension regions with a low sheet resistance.1 Among the other processes nonmelt laser annealing has gained attention as a means of achieving these requirements by its short process time and high annealing temperature, and hence low thermal budget, resulting in high dopant solubility.2-5 A problem exists with creating highly active profiles when boron is implanted in conjunction with a preamorphizing germanium implant; deactivation occurs during postactivation thermal processes. [6][7][8][9][10][11][12] This deactivation is thought to be driven by the release of silicon interstitials from end-of-range ͑EOR͒ defects that evolve through nonconservative Ostwald ripening during annealing. 13 The interstitials flow towards the surface and decorate the boron profile, producing boron interstitial clusters. [14][15][16][17] In this letter, multiple laser scan annealing at 1150°C followed by isochronal rapid thermal postannealing at lower temperatures is used to investigate the role of end-of-range defects in the redistribution and deactivation of ultrashallow B profiles in preamorphized and nonmelt laser-annealed silicon.N-type ͑100͒ Czochralski-silicon wafers were preamorphized with 5 keV Ge + to a dose of 1 ϫ 10 15 cm −2 producing a surface amorphous layer to a depth of ϳ15 nm. 500 eV B + was implanted into the amorphous layer to a dose of 1 ϫ 10 15 cm −2 . Both implants were made using an Applied Materials Quantum X implanter. The wafers were exposed to a scanning diode laser source operated under nonmelting conditions, which was used to anneal three strips across the wafers, corresponding to one, five, or ten scans at a temperature of 1150°C. By using multiple laser scans to anneal the wafer, it allows a study of defect evolution as a function of increasing the thermal budget. The amorphous layer regrew by solid phase epitaxial regrowth during the annealing. Samples were taken from these strips and annealed in dry N 2 for 60 s at temperatures ranging from 700 to 1000°C using a Process Products Corporation rapid thermal annealing system operating with a 50°C/s heating ramp rate. The van der Pauw technique was use...

Deep neural networks (DNNs) have a wide range of applications, and software employing them must be thoroughly tested, especially in safety-critical domains. However, traditional software test coverage metrics cannot be applied directly to DNNs. In this paper, inspired by the MC/DC coverage criterion, we propose a family of four novel test coverage criteria that are tailored to structural features of DNNs and their semantics. We validate the criteria by demonstrating that test inputs that are generated with guidance by our proposed coverage criteria are able to capture undesired behaviours in a DNN. Test cases are generated using a symbolic approach and a gradient-based heuristic search. By comparing them with existing methods, we show that our criteria achieve a balance between their ability to find bugs (proxied using adversarial examples and correlation with functional coverage) and the computational cost of test input generation. Our experiments are conducted on state-of-the-art DNNs obtained using popular open source datasets, including MNIST, CIFAR-10 and ImageNet.

Abstract. Much research work has been done on formalizing UML diagrams, but less has focused on using this formalization to analyze the dynamic behaviours between formalized components. In this paper we propose using a subset of fUML (Foundational Subset for Executable UML) as a semi-formal language, and formalizing it to the process algebraic specification language CSP, to make use of FDR as a model checker. Our formalization includes modelling the asynchronous communication framework used within fUML. This allows different interpretations of the communications model to be evaluated. To illustrate the approach, we use the modelling of the Tokeneer ID Station specifications into fUML, and formalize them in CSP to check if the model is deadlock free.

No abstract

Recurrent neural networks (RNNs) have been applied to a broad range of applications including natural language processing, drug discovery, and video recognition. However, their vulnerability to input perturbation is also exposed. Aligning with a view from software defect detection, this paper aims to develop a coverage guided testing approach to systematically exploit the internal behaviour of RNNs, with high possibility of detecting defects. Technically, the long short term memory network (LSTM), a major class of RNN, is thoroughly studied. A family of three test metrics are designed to quantify not only the values but also the temporal relations (including both step-wise and bounded-length) exhibited when LSTM processing inputs. A genetic algorithm is applied to efficiently generate test cases. Based on these, we develop a tool TESTRNN, and extensively evaluate TESTRNN on a set of LSTM benchmarks. Experiments confirm that TESTRNN has several advantages over the state-of-art tool DeepStellar and attack-based defect detection methods, owing to its working with finer temporal semantics and the consideration of the naturalness of input perturbation. Furthermore, TESTRNN enables meaningful information to be collected and exhibited for users to understand the testing results, which is an important step towards interpretable neural network testing.

No abstract