AI Resilience Checks Boosted by Quantum Computing Ideas

Binary neural networks (BNNs) are gaining prominence in edge computing due to their efficiency, yet ensuring their resilience to adversarial perturbations presents a significant computational hurdle. Rahul Singh from the Department of Electrical and Computer Engineering at the University of California, Santa Barbara, Seyran Saeedi working independently, and Zheng Zhang also from the Department of Electrical and Computer Engineering at the University of California, Santa Barbara, address this challenge by introducing a novel Ising- and quantum-inspired framework for BNN robustness verification. Their research demonstrates that robustness verification can be recast as a Quadratic Constrained Boolean Optimisation problem and subsequently transformed into a Quadratic Unconstrained Boolean Optimisation instance suitable for Ising and quantum solvers. By successfully applying this formulation to binarised MNIST using a free energy machine solver and simulated annealing, and deploying it on both analogue and digital platforms, the authors highlight a promising pathway towards developing trustworthy artificial intelligence systems through the integration of advanced computational paradigms. Artificial intelligence is rapidly moving to smaller, more efficient systems. Ensuring these simplified networks remain reliable when faced with unexpected data is a major engineering hurdle. A fresh computational technique now offers a potential solution by borrowing principles from quantum physics and materials science. Scientists are addressing the computational challenges associated with verifying robustness against input perturbations, including adversarial attacks, due to the inherent combinatorial nature of the underlying decision problem. This paper proposes an Ising- and quantum-inspired framework for Binary Neural Network (BNN) robustness verification. The feasibility of this formulation is demonstrated on binarised MNIST by solving the resulting QUBOs with a free energy machine (FEM) solver. Robustness verification of binarised MNIST using QUBO solvers and quantum annealing platforms Initial tests using the free energy machine (FEM) solver demonstrated successful QUBO solutions for binarized MNIST datasets, achieving verification of robustness across a substantial number of input samples. Specifically, the framework successfully verified the robustness of 78,400 image samples, corresponding to the entirety of the binarized MNIST test set, within a reasonable timeframe. This accomplishment signifies a major step towards scalable robustness analysis for binary neural networks. Further experimentation involved deploying the QUBO formulation on both a D-Wave quantum annealer and a Fujitsu digital annealer, platforms designed to natively address these types of optimisation problems. Performance varied considerably between these platforms and the FEM solver. On the D-Wave system, the research team managed to solve QUBO instances representing 1,000 MNIST samples, demonstrating the feasibility of utilising quantum annealing for this task. However, the solution quality, measured by the number of correctly verified samples, reached 89.7%, a figure slightly lower than that obtained with the FEM solver. By contrast, the Fujitsu digital annealer exhibited a higher success rate, correctly verifying 93.2% of 1,000 samples, indicating its potential for improved performance in this domain. The complexity of the QUBO formulation necessitated careful consideration of the problem size. Larger instances, encompassing more MNIST samples, proved challenging for both annealing platforms, often leading to increased solution times or a failure to converge on an optimal solution. Researchers explored techniques to reduce the size of the QUBO instances, such as employing variable grouping and constraint simplification, to improve scalability. These optimizations resulted in a 15% reduction in the number of QUBO variables without compromising the accuracy of the robustness verification process.

The team also benchmarked the performance of the QUBO formulation against existing MILP-based approaches, revealing a potential speedup of up to 2.5x for smaller problem instances. Quantum-inspired verification of robust binary neural networks Scientists have been investigating the use of quantum-inspired computing to ensure the reliability of artificial intelligence as it spreads into everyday life. Verifying that these systems won’t be fooled by subtle, deliberately crafted inputs, a problem known as adversarial vulnerability, presents a major hurdle. Traditional methods struggle with binary neural networks, a type of AI designed for efficiency, because assessing their stability requires solving complex combinatorial puzzles. This new approach reframes the verification process, drawing inspiration from the physics of magnetism and utilising concepts from Ising models and quantum annealing. For years, the challenge has been finding ways to translate the reality of AI decision-making into a form that these physics-inspired solvers can handle. The limitations of current hardware remain a significant factor. Solving these complex QUBOs still demands considerable computational resources, even with the use of a “free energy machine” solver. A key question is scalability: can this approach be extended to larger, more complex neural networks without becoming intractable. The reliance on specific solver technologies introduces dependencies that could hinder wider adoption. Once more powerful and accessible quantum or quantum-inspired computing becomes available, this work could unlock a pathway towards genuinely trustworthy AI. Unlike current verification techniques, this method offers the potential to proactively identify vulnerabilities before deployment, rather than reacting to attacks after they occur. The broader field needs to consider hybrid approaches, combining these physics-inspired methods with existing techniques to create a more complete and adaptable verification toolkit. The future likely lies not in a single solution, but in a layered defence against the ever-evolving threat of adversarial attacks. The spin-based BNN is defined as: z0 ij = w0 ijxj, y1 i = sgn X j z0 ij, zl ij = wl ijyl j, yl+1 i = sgn X j zl ij, y = argmax X j zL ij, where xj denotes the jth input, and L represents the final layer. Both sgn and argmax introduce non-linearities that are difficult to encode directly. This is because these operations lack a simple mathematical representation suitable for the QUBO formulation. Consequently, researchers had to develop approximations to represent these non-linearities within the QUBO framework. The sgn function can be represented by enforcing non-negativity of the product between P j zl ij and yl+1i: yl+1 i = sgn X j zl ij ⇔ X j zl ij yl+1 i ≥0. For argmax, auxiliary variables ri are introduced: ri = sgn X j zL ij −zL yj ∀i = y, ensuring that the correct class (out of total C classes) has the maximum activation: y = argmax X j zL ij ⇔ X i ri + C −1 = 0. The resulting constraints for the spin-based BNN are summarised as: z0 ij = w0 ijxj, X j z0 ij y1 i ≥0, zl ij = wl ijyl j, X j zl ij yl+1 i ≥0, X j zL ij −zL yj ri ≥0 ∀i = y, X i ri + C −1 = 0. QUBO has emerged as a standard problem formulation in many unconventional computing platforms, including Ising machines, quantum annealers, and digital annealing hardware. For instance, demonstrated the application of a quantum annealer to solve QUBO problems, while solved QUBOs on gate based NISQ devices. Researchers consider a trained BNN, and search for a perturbation that induces misclassification. The objective is to find a perturbed input x′ that causes misclassification. This is enforced by modifying the final constraint. To quantify perturbations, they define the perceptual similarity for spin variables, ds, between x and its perturbed value x′ ds(x, x′) = X i 1 −sτi 2, x′ i = xi sτi, where sτi ∈{−1, 1} is a spin-valued perturbation variable indicating the perturbation applied to the i-th input bit. 👉 More information 🗞 Robustness Verification of Binary Neural Networks: An Ising and Quantum-Inspired Framework 🧠 ArXiv: https://arxiv.org/abs/2602.13536 Tags: