HeLayers: A Tile Tensors Framework for Large Neural Networks on Encrypted Data

Aharoni, Ehud; Adir, Allon; Baruch, Moran; Drucker, Nir; Ezov, Gilad; Farkash, Ariel; Greenberg, Lev; Masalha, Ramy; Moshkowich, Guy; Murik, Dov; Shaul, Hayim; Soceanu, Omri

doi:10.48550/arxiv.2011.01805

Cited by 2 publications

(7 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Figure 2 demonstrates three examples for tiling a matrix M [5,6]. In Figure 2c, for example, the shape of the external tensor is [3,2], and the tile shape is [2, 4].…”

Section: Tiling Basicsmentioning

confidence: 99%

“…It works by summing over the external tensor along the i'th dimension, then by summing inside each tile along the i'th dimension. In an HE environment, the latter summation requires using the rotate-and-sum algorithm [3]. Generally, the sum operator reduces the i'th dimension and the resulting tile tensor shape changes to 1?…”

Section: Operatorsmentioning

confidence: 99%

“…Figures 4a and 4b show a matrix M [6,6] packed in the tile tensor T M [ 6∼ 2 , 6∼ 4 ]. Here, the tile shape is [2, 4] and the external tensor shape is [3,2]. Every tile contains a 2 × 4 sub-matrix, but instead of being contiguous, it is a set of elements spaced evenly in the matrix.…”

Section: Convolution With Interleaved Dimensionsmentioning

confidence: 99%

“…Note further that the external sizes e 1 = w I t 1 and e 2 = h I t 2 of the tile tensor T I remain the same in T O , and they may be larger than those actually required to hold the tensor The convolution is computed similarly to the description in Section 5.1, multiplying tiles of T I with the appropriate tiles of T F . The result is a tile tensor of shape , where the log t 3 ) factor comes from the sum-and-rotate algorithm [3].…”

Section: Proof See Appendix E □mentioning

confidence: 99%

“…Let E 1 = sum(E, i). E ′ is computed from E 1 by summing over the dimension i of every tile L of E 1 using the rotate-and-sum algorithms [3]. S ′ is identical to S except n ′ i = 1, and if ∀ j <i t j = 1 and t i is a power of 2, then r ′ i = t i , else u ′ i = (t i > 1).…”

Section: A Experiments Setupmentioning

confidence: 99%

See 4 more Smart Citations

HeLayers: A Tile Tensors Framework for Large Neural Networks on Encrypted Data

Aharoni

Adir

Baruch

et al. 2023

PoPETs

View full text Add to dashboard Cite

Privacy-preserving solutions enable companies to offload confidential data to third-party services while fulfilling their government regulations. To accomplish this, they leverage various cryptographic techniques such as Homomorphic Encryption (HE), which allows performing computation on encrypted data. Most HE schemes work in a SIMD fashion, and the data packing method can dramatically affect the running time and memory costs. Finding a packing method that leads to an optimal performant implementation is a hard task. We present a simple and intuitive framework that abstracts the packing decision for the user. We explain its underlying data structures and optimizer, and propose a novel algorithm for performing 2D convolution operations. We used this framework to implement an inference operation over an encrypted HE-friendly AlexNet neural network with large inputs, which runs in around five minutes, several orders of magnitude faster than other state-of-the-art non-interactive HE solutions.

show abstract

“…Figure 2 demonstrates three examples for tiling a matrix M [5,6]. In Figure 2c, for example, the shape of the external tensor is [3,2], and the tile shape is [2, 4].…”

Section: Tiling Basicsmentioning

confidence: 99%

Section: Operatorsmentioning

confidence: 99%