Search for: All records

Abstract Deep Neural Networks (DNNs) are increasingly deployed in critical applications, where ensuring their safety and robustness is paramount. We present$$_\text {CAV25}$$ ${}_{\mathrm{CAV}25}$, a high-performance DNN verification tool that uses the DPLL(T) framework and supports a wide-range of network architectures and activation functions. Since its debut in VNN-COMP’23, in which it achieved the New Participant Award and ranked 4th overall,$$_\text {CAV25}$$ ${}_{\mathrm{CAV}25}$has advanced significantly, achieving second place in VNN-COMP’24. This paper presents and evaluates the latest development of$$_\text {CAV25}$$ ${}_{\mathrm{CAV}25}$, focusing on the versatility, ease of use, and competitive performance of the tool.$$_\text {CAV25}$$ ${}_{\mathrm{CAV}25}$is available at:https://github.com/dynaroars/neuralsat. 

Deep Neural Networks (DNN) have emerged as an effective approach to tackling real-world problems. However, like human-written software, DNNs are susceptible to bugs and attacks. This has generated significant interests in developing effective and scalable DNN verification techniques and tools. Recent developments in DNN verification have highlighted the potential of constraint-solving approaches that combine abstraction techniques with SAT solving. Abstraction approaches are effective at precisely encode neuron behavior when it is linear, but they lead to overapproximation and combinatorial scaling when behavior is non-linear. SAT approaches in DNN verification have incorporated standard DPLL techniques, but have overlooked important optimizations found in modern SAT solvers that help them scale on industrial benchmarks. In this paper, we present VeriStable, a novel extension of recently proposed DPLL-based constraint DNN verification approach. VeriStable leverages the insight that while neuron behavior may be non-linear across the entire DNN input space, at intermediate states computed during verification many neurons may be constrained to have linear behavior – these neurons are stable. Efficiently detecting stable neurons reduces combinatorial complexity without compromising the precision of abstractions. Moreover, the structure of clauses arising in DNN verification problems shares important characteristics with industrial SAT benchmarks. We adapt and incorporate multi-threading and restart optimizations targeting those characteristics to further optimize DPLL-based DNN verification. We evaluate the effectiveness of VeriStable across a range of challenging benchmarks including fully- connected feedforward networks (FNNs), convolutional neural networks (CNNs) and residual networks (ResNets) applied to the standard MNIST and CIFAR datasets. Preliminary results show that VeriStable is competitive and outperforms state-of-the-art DNN verification tools, including 𝛼-𝛽-CROWN and MN-BaB, the first and second performers of the VNN-COMP, respectively. 

Deep Neural Networks (DNN) have emerged as an effective approach to tackling real-world problems. However, like human-written software, DNNs are susceptible to bugs and attacks. This has generated significant interest in developing effective and scalable DNN verification techniques and tools. Recent developments in DNN verification have highlighted the potential of constraint-solving approaches that combine abstraction techniques with SAT solving. Abstraction approaches are effective at precisely encoding neuron behavior when it is linear, but they lead to overapproximation and combinatorial scaling when behavior is non-linear. SAT approaches in DNN verification have incorporated standard DPLL techniques, but have overlooked important optimizations found in modern SAT solvers that help them scale on industrial benchmarks. In this paper, we present VeriStable, a novel extension of the recently proposed DPLL-based constraint DNN verification approach. VeriStable leverages the insight that while neuron behavior may be non-linear across the entire DNN input space, at intermediate states computed during verification many neurons may be constrained to have linear behavior – these neurons are stable. Efficiently detecting stable neurons reduces combinatorial complexity without compromising the precision of abstractions. Moreover, the structure of clauses arising in DNN verification problems shares important characteristics with industrial SAT benchmarks. We adapt and incorporate multi-threading and restart optimizations targeting those characteristics to further optimize DPLL-based DNN verification. We evaluate the effectiveness of VeriStable across a range of challenging benchmarks including fully- connected feedforward networks (FNNs), convolutional neural networks (CNNs) and residual networks (ResNets) applied to the standard MNIST and CIFAR datasets. Preliminary results show that VeriStable is competitive and outperforms state-of-the-art DNN verification tools, including α-β-CROWN and MN-BaB, the first and second performers of the VNN-COMP, respectively. 

A key challenge in program synthesis is synthesizing programs that use libraries, which most real-world software does. The current state of the art is to model libraries with mock library implementations that perform the same function in a simpler way. However, mocks may still be large and complex, and must include many implementation details, both of which could limit synthesis performance. To address this problem, we introduce JLibSketch, a Java program synthesis tool that allows library behavior to be described with algebraic specifications, which are rewrite rules for sequences of method calls, e.g., encryption followed by decryption (with the same key) is the identity. JLibSketch implements rewrite rules by compiling JLibSketch problems into problems for the Sketch program synthesis tool. More specifically, after compilation, library calls are represented by abstract data types (ADTs), and rewrite rules manipulate those ADTs. We formalize compilation and prove it sound and complete if the rewrite rules are ordered and non-unifiable. We evaluated JLibSketch by using it to synthesize nine programs that use libraries from three domains: data structures, cryptography, and file systems. We found that algebraic specifications are, on average, about half the size of mocks. We also found that algebraic specifications perform better than mocks on seven of the nine programs, sometimes significantly so, and perform equally well on the last two programs. Thus, we believe that JLibSketch takes an important step toward synthesis of programs that use libraries.