Neural networks are increasingly employed in safety-critical domains. This has prompted interest in verifying or certifying logically encoded properties of neural networks. Prior work has largely focused on checking existential properties, wherein the goal is to check whether there exists any input that violates a given property of interest. However, neural network training is a stochastic process, and many questions arising in their analysis require probabilistic and quantitative reasoning, i.e., estimating how many inputs satisfy a given property. To this end, our paper proposes a novel and principled framework to quantitative verification of logical properties specified over neural networks. Our framework is the first to provide PAC-style soundness guarantees, in that its quantitative estimates are within a controllable and bounded error from the true count. We instantiate our algorithmic framework by building a prototype tool called NPAQ that enables checking rich properties over binarized neural networks. We show how emerging security analyses can utilize our framework in 3 concrete point applications: quantifying robustness to adversarial inputs, efficacy of trojan attacks, and fairness/bias of given neural networks.
Dynamic binary taint analysis has wide applications in the security analysis of commercial-off-the-shelf (COTS) binaries. One of the key challenges in dynamic binary analysis is to specify the taint rules that capture how taint information propagates for each instruction on an architecture. Most of the existing solutions specify taint rules using a deductive approach by summarizing the rules manually after analyzing the instruction semantics. Intuitively, taint propagation reflects on how an instruction input affects its output, and thus can be observed from instruction executions. In this work, we propose an inductive method for taint propagation and develop a universal taint tracking engine that is architecture-agnostic. Our taint engine, TAINTINDUCE, can learn taint rules with minimal architectural knowledge by observing the execution behavior of instructions. To measure its correctness and guide taint rule generation, we define the precise notion of soundness for bit-level taint tracking in this novel setup. In our evaluation, we show that TAINTINDUCE automatically learns rules for 4 widely used architectures: x86, x64, AArch64, and MIPS-I. It can detect vulnerabilities for 24 CVEs in 15 applications on both Linux and Windows over millions of instructions and is comparable with other mature existing tools (TEMU [51], libdft [32], Triton [42]). TAINTINDUCE can be used as a stand-alone taint engine or be used to complement existing taint engines for unhandled instructions. Further, it can be used as a cross-referencing tool to uncover bugs in taint engines, emulation implementations and ISA documentations.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.