arXiv subjects

Tomer Pelleg

Publications and source records attributed to Tomer Pelleg.

2 recordsLinked to original sources

arXiv2025

Testing forbidden order-pattern properties on hypergrids

We study testing $\pi$-freeness of functions $f:[n]^d\to\mathbb{R}$, where $f$ is $\pi$-free if there there are no $k$ indices $x_1\prec\cdots\prec x_k\in [n]^d$ such that $f(x_i) 2$. We initiate a systematic study of pattern freeness on higher-dimensional grids. For $d=2$ and all permutations of size $k=3$, we design an adaptive one-sided tester with query complexity $O(n^{4/5+o(1)})$. We also prove general lower bounds for $k=3$: every nonadaptive tester requires $\Omega(n)$ queries, and every adaptive tester requires $\Omega(\sqrt{n})$ queries, yielding the first super-logarithmic lower bounds for $\pi$-freeness. For the monotone patterns $\pi=(1,2,3)$ and $(3,2,1)$, we present a nonadaptive tester with polylogarithmic query complexity, giving an exponential separation between monotone and nonmonotone patterns (unlike the one-dimensional case). A key ingredient in our $\pi$-freeness testers is new erasure-resilient ($\delta$-ER) $\epsilon$-testers for monotonicity over $[n]^d$ with query complexity $O(\log^{O(d)}n/(\epsilon(1-\delta)))$, where $0<\delta<1$ is an upper bound on the fraction of erasures. Prior ER testers worked only for $\delta=O(\epsilon/d)$. Our nonadaptive monotonicity tester is nearly optimal via a matching lower bound due to Pallavoor, Raskhodnikova, and Waingarten (Random Struct. Algorithms, 2022). Finally, we show that current techniques cannot yield sublinear-query testers for patterns of length $4$ even on two-dimensional hypergrids.

cs.DS↗

arXiv2024

Efficient Pruning for Machine Learning Under Homomorphic Encryption

Privacy-preserving machine learning (PPML) solutions are gaining widespread popularity. Among these, many rely on homomorphic encryption (HE) that offers confidentiality of the model and the data, but at the cost of large latency and memory requirements. Pruning neural network (NN) parameters improves latency and memory in plaintext ML but has little impact if directly applied to HE-based PPML. We introduce a framework called HE-PEx that comprises new pruning methods, on top of a packing technique called tile tensors, for reducing the latency and memory of PPML inference. HE-PEx uses permutations to prune additional ciphertexts, and expansion to recover inference loss. We demonstrate the effectiveness of our methods for pruning fully-connected and convolutional layers in NNs on PPML tasks, namely, image compression, denoising, and classification, with autoencoders, multilayer perceptrons (MLPs) and convolutional neural networks (CNNs). We implement and deploy our networks atop a framework called HElayers, which shows a 10-35% improvement in inference speed and a 17-35% decrease in memory requirement over the unpruned network, corresponding to 33-65% fewer ciphertexts, within a 2.5% degradation in inference accuracy over the unpruned network. Compared to the state-of-the-art pruning technique for PPML, our techniques generate networks with 70% fewer ciphertexts, on average, for the same degradation limit.

cs.CR↗