More Like this

1. SCANN: Side Channel Analysis of Spiking Neural Networks

   https://doi.org/10.3390/cryptography7020017  

   Nagarajan, Karthikeyan; Roy, Rupshali; Topaloglu, Rasit Onur; Kannan, Sachhidh; Ghosh, Swaroop (June 2023, Cryptography)

   Spiking neural networks (SNNs) are quickly gaining traction as a viable alternative to deep neural networks (DNNs). Compared to DNNs, SNNs are computationally more powerful and energy efficient. The design metrics (synaptic weights, membrane threshold, etc.) chosen for such SNN architectures are often proprietary and constitute confidential intellectual property (IP). Our study indicates that SNN architectures implemented using conventional analog neurons are susceptible to side channel attack (SCA). Unlike the conventional SCAs that are aimed to leak private keys from cryptographic implementations, SCANN (SCA̲ of spiking n̲eural n̲etworks) can reveal the sensitive IP implemented within the SNN through the power side channel. We demonstrate eight unique SCANN attacks by taking a common analog neuron (axon hillock neuron) as the test case. We chose this particular model since it is biologically plausible and is hence a good fit for SNNs. Simulation results indicate that different synaptic weights, neurons/layer, neuron membrane thresholds, and neuron capacitor sizes (which are the building blocks of SNN) yield distinct power and spike timing signatures, making them vulnerable to SCA. We show that an adversary can use templates (using foundry-calibrated simulations or fabricating known design parameters in test chips) and analysis to identify the specifications of the implemented SNN. 

   more » « less
2. A Greedy Algorithm for Quantizing Neural Networks

   Lybrand, Eric; Saab, Rayan (January 2021, Journal of machine learning research)

   We propose a new computationally efficient method for quantizing the weights of pre- trained neural networks that is general enough to handle both multi-layer perceptrons and convolutional neural networks. Our method deterministically quantizes layers in an iterative fashion with no complicated re-training required. Specifically, we quantize each neuron, or hidden unit, using a greedy path-following algorithm. This simple algorithm is equivalent to running a dynamical system, which we prove is stable for quantizing a single-layer neural network (or, alternatively, for quantizing the first layer of a multi-layer network) when the training data are Gaussian. We show that under these assumptions, the quantization error decays with the width of the layer, i.e., its level of over-parametrization. We provide numerical experiments, on multi-layer networks, to illustrate the performance of our methods on MNIST and CIFAR10 data, as well as for quantizing the VGG16 network using ImageNet data. 

   more » « less
3. Sensing and actuation are intricately connected in soft robotics, where contact may change actuator mechanics and robot behavior. To improve soft robotic control and performance, proprioception and contact sensors are needed to report robot state without altering actuation mechanics or introducing bulky, rigid components. For bioinspired McKibben-style fluidic actuators, prior work in sensing has focused on sensing the strain of the actuator by embedding sensors in the actuator bladder during fabrication, or by adhering sensors to the actuator surface after fabrication. However, material property mismatches between sensors and actuators can impede actuator performance, and many soft sensors available for use with fluidic actuators rely on costly or labor-intensive fabrication methods. Here, we demonstrate a low-cost and easy-to manufacture-tubular liquid metal strain sensor for use with soft actuators that can be used to detect actuator strain and contact between the actuator and external objects. The sensor is flexible, can be fabricated with commercial-off-the-shelf components, and can be easily integrated with existing soft actuators to supplement sensing, regardless of actuator shape or size. Furthermore, the soft tubular strain sensor exhibits low hysteresis and high sensitivity. The approach presented in this work provides a low-cost, soft sensing solution for broad application in soft robotics.

   Jackson, Clayton; Chardon, Matthieu; Wang, Y Curtis; Rudi, Johann; Tresch, Matthew; Heckman, Charles J; Quinn, Roger D (August 2023, Springer, Cham.)

   This work explored synaptic strengths in a computational neuroscience model of a controller for the hip joint of a rat which consists of Ia interneurons, Renshaw cells, and the associated motor neurons. This circuit has been referred to as the Canonical Motor Microcircuit (CMM). It is thought that the CMM acts to modulate motor neuron activity at the output stage. We first created a biomechanical model of a rat hindlimb consisting of a pelvis, femur, shin, foot, and flexor-extensor muscle pairs modeled with a Hill muscle model. We then modeled the CMM using non-spiking leaky-integrator neural models connected with conductance-based synapses. To tune the parameters in the network, we implemented an automated approach for parameter search using the Markov chain Monte Carlo (MCMC) method to solve a parameter estimation problem in a Bayesian inference framework. As opposed to traditional optimization techniques, the MCMC method identifies probability densities over the multidimensional space of parameters. This allows us to see a range of likely parameters that produce model outcomes consistent with animal data, determine if the distribution of likely parameters is uni- or multi-modal, as well as evaluate the significance and sensitivity of each parameter. This approach will allow for further analysis of the circuit, specifically, the function and significance of Ia feedback and Renshaw cells. 

   more » « less
4. Surrogate Bayesian Networks for Approximating Evolutionary Games

   Hsiao, Vincent; Nau, Dana S; Dechter, Rina (May 2024, PMLR)

   Dasgupta, Sanjoy; Mandt, Stephan; Li, Yingzhen (Ed.)

   Spatial evolutionary games are used to model large systems of interacting agents. In earlier work, a method was developed using Bayesian Networks to approximate the population dynamics in these games. One advantage of that approach is that one can smoothly adjust the size of the network to get more accurate approximations. However, scaling the method up can be intractable if the number of strategies in the evolutionary game increases. In this paper, we propose a new method for computing more accurate approximations by using surrogate Bayesian Networks. Instead of doing inference on larger networks directly, we do it on a much smaller surrogate network extended with parameters that exploit the symmetry inherent to the domain. We learn the parameters on the surrogate network using KL-divergence as the loss function. We illustrate the value of this method empirically through a comparison on several evolutionary games. 

   more » « less
5. Surrogate {B}ayesian Networks for Approximating Evolutionary Games

   Hsiao, Vincent; Nau, Dana; Pezeshki, Bobak; Dechter, Rina (May 2025, PMLR)

   Dasgupta, Sanjoy; Mandt, Stephan; Li, Yingzhen (Ed.)

   Spatial evolutionary games are used to model large systems of interacting agents. In earlier work, a method was developed using Bayesian Networks to approximate the population dynamics in these games. One advantage of that approach is that one can smoothly adjust the size of the network to get more accurate approximations. However, scaling the method up can be intractable if the number of strategies in the evolutionary game increases. In this paper, we propose a new method for computing more accurate approximations by using surrogate Bayesian Networks. Instead of doing inference on larger networks directly, we do it on a much smaller surrogate network extended with parameters that exploit the symmetry inherent to the domain. We learn the parameters on the surrogate network using KL-divergence as the loss function. We illustrate the value of this method empirically through a comparison on several evolutionary games. 

   more » « less