1 4 : 3 5 P M 1 5 : 1 0 P M C r y p Te n : A M L f r a m e w o - PowerPoint PPT Presentation

1 4 : 3 5 P M — 1 5 : 1 0 P M C r y p Te n : A M L f r a m e w o r k b a s e d o n s e c u r e c o m p u t i n g t e c h n i q u e s Shubho Sengupta Facebook

import crypten import torch crypten.init() # sets up # communication x = torch.tensor([1.0, 2.0, 3.0]) x_enc = crypten.cryptensor(x) # encrypts tensor Why is CrypTen the way it is? x_dec = x_enc.get_plain_text() # decrypts tensor assert torch.all_close(x_dec, x) # this passes! y_enc = crypten.cryptensor([2.0, 3.0, 4.0]) xy_enc = x_enc + y_enc # adds encrypted # tensors xy_dec = xy_enc.get_plain_text() assert torch.all_close(xy_dec, x + y) # this passes!

Expose the Machine Learning Community to Secure Computing Techniques

ML Centric Interface

ML Centric Interface Explainable Performance

ML Centric Interface Explainable Performance Debuggability

ML Centric Interface Explainable Performance Debuggability Interoperability

Honest but curious ML Centric Interface Explainable Performance Debuggability Interoperability

What makes an ML framework?

Tensor What makes an ML framework?

Tensor Operators What makes an ML framework?

Tensor Operators What makes an ML framework? Computation graph

Tensor Operators What makes an ML framework? Computation graph Gradients

Mul

Mul

Mul

Mul BMul

AccumGrad AccumGrad Mul BMul

AccumGrad AccumGrad Mul BMul

CrypTensor

CrypTensor MPCTensor

CrypTensor MPCTensor ArithmeticShared Tensor

CrypTensor MPCTensor ArithmeticShared BinaryShared Tensor Tensor

CrypTensor MPCTensor ArithmeticShared BinaryShared Tensor Tensor LongTensor

CrypTensor MPCTensor ArithmeticShared BinaryShared Tensor Tensor LongTensor

CrypTensor MPCTensor ArithmeticShared BinaryShared Tensor Tensor LongTensor PyTorch

CrypTensor Interface filters up MPCTensor ArithmeticShared BinaryShared Tensor Tensor LongTensor PyTorch

CrypTensor Interface filters up MPCTensor Performance is linked ArithmeticShared BinaryShared Tensor Tensor LongTensor PyTorch

CrypTensor Interface filters up MPCTensor Performance is linked ArithmeticShared BinaryShared Communication libraries Tensor Tensor LongTensor PyTorch

What operations do we need for Machine Learning training?

matmul() , conv2d() What operations do we need for Machine Learning training?

matmul() , conv2d() log() , exp() What operations do we need for Machine Learning training?

matmul() , conv2d() log() , exp() What operations do we div() need for Machine Learning training?

matmul() , conv2d() log() , exp() What operations do we div() need for Machine Learning training? pow() , sqrt()

matmul() , conv2d() log() , exp() What operations do we div() need for Machine Learning training? pow() , sqrt() relu() , max() , argmax()

CrypTensor MPCTensor ArithmeticShared BinaryShared Tensor Tensor LongTensor PyTorch

CrypTensor

CrypTensor

AutogradCrypTensor CrypTensor

Module AutogradCrypTensor CrypTensor

Module Compatibility AutogradCrypTensor CrypTensor

Module Compatibility AutogradCrypTensor Use pre-trained models CrypTensor

Module Compatibility AutogradCrypTensor Use pre-trained models Train models from scratch CrypTensor

AllReduce

Open to all AllReduce

Open to all AllReduce Reduce

Open to all AllReduce Open to one Reduce

Open to all AllReduce Open to one Reduce Broadcast

Open to all AllReduce Open to one Reduce Trusted dealer to parties Broadcast

PyTorch CrypTen

PyTorch CrypTen import torch x = torch.tensor([1.0, 2.0, 3.0]) y = torch.tensor([2.0, 3.0, 4.0]) xy = x + y

PyTorch CrypTen import crypten import torch crypten.init() # sets up # communication x_enc = crypten.cryptensor([1.0, 2.0, 3.0]) x = torch.tensor([1.0, 2.0, 3.0]) y_enc = crypten.cryptensor([2.0, 3.0, 4.0]) y = torch.tensor([2.0, 3.0, 4.0]) xy_enc = x_enc + y_enc xy = x + y

PyTorch CrypTen pt = torch.tensor([1.0, 2.0, 3.0]) pt = torch.tensor([1.0, 2.0, 3.0]) ct = AutogradCrypTensor(crypten.cryptensor(pt)) ct_loss = ct.cross_entropy() pt_loss = pt.cross_entropy() ct_loss.backward() pt_loss.backward()

How do we work with real model and datasets?

import crypten import torchvision.datasets as datasets import torchvision.models as models crypten.init() How do we work with real model and datasets?

import crypten import torchvision.datasets as datasets import torchvision.models as models crypten.init() # download and set up ImageNet dataset: transform = transforms.Compose([ transforms.Resize(256), How do we work with transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize(mean=[0.48, 0.45, 0.41], std=[0.23, 0.22, 0.23]), real model and datasets? ]) dataset = datasets.ImageNet(imagenet_folder, transform=transform)

import crypten import torchvision.datasets as datasets import torchvision.models as models crypten.init() # download and set up ImageNet dataset: transform = transforms.Compose([ transforms.Resize(256), How do we work with transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize(mean=[0.48, 0.45, 0.41], std=[0.23, 0.22, 0.23]), real model and datasets? ]) dataset = datasets.ImageNet(imagenet_folder, transform=transform) # download pre-trained ResNet-18 model and encrypt it: model = models.resnet18(pretrained=True) encrypted_model = crypten.nn.from_pytorch(model, dataset[0])

import crypten import torchvision.datasets as datasets import torchvision.models as models crypten.init() # download and set up ImageNet dataset: transform = transforms.Compose([ transforms.Resize(256), How do we work with transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize(mean=[0.48, 0.45, 0.41], std=[0.23, 0.22, 0.23]), real model and datasets? ]) dataset = datasets.ImageNet(imagenet_folder, transform=transform) # download pre-trained ResNet-18 model and encrypt it: model = models.resnet18(pretrained=True) encrypted_model = crypten.nn.from_pytorch(model, dataset[0]) # do inference on encrypted images with encrypted model: encrypted_image = crypten.cryptensor(dataset[1]) encrypted_output = encrypted_model(encrypted_image)

import crypten import torchvision.datasets as datasets import torchvision.models as models crypten.init() # download and set up ImageNet dataset: transform = transforms.Compose([ transforms.Resize(256), How do we work with transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize(mean=[0.48, 0.45, 0.41], std=[0.23, 0.22, 0.23]), real model and datasets? ]) dataset = datasets.ImageNet(imagenet_folder, transform=transform) # download pre-trained ResNet-18 model and encrypt it: model = models.resnet18(pretrained=True) encrypted_model = crypten.nn.from_pytorch(model, dataset[0]) # do inference on encrypted images with encrypted model: encrypted_image = crypten.cryptensor(dataset[1]) encrypted_output = encrypted_model(encrypted_image) output = encrypted_output.get_plain_text() # this works!

Improving Performance

Improving Performance More Secure TTP

Support for other Improving Performance techniques? More Secure TTP

Support for other Improving Performance techniques? More Secure TTP Quantifying information leakage?

ILSVRC Challenge

ILSVRC Challenge Train on a million images

ILSVRC Challenge Train on a million images Classify with high accuracy

ILSVRC Challenge Train on a million images Classify with high accuracy Done in minutes

ILSVRC Challenge Question? Train on a million images Classify with high accuracy Done in minutes

ILSVRC Challenge Question? Train on a million images Train on a million encrypted images Classify with high accuracy Done in minutes

ILSVRC Challenge Question? Train on a million images Train on a million encrypted images Classify with high accuracy Classify with high accuracy Done in minutes

ILSVRC Challenge Question? Train on a million images Train on a million encrypted images Classify with high accuracy Classify with high accuracy Done in minutes Done in a week

CrypTen Team at Facebook

Awni CrypTen Team at Facebook

Awni Brian CrypTen Team at Facebook

Awni Brian Laurens CrypTen Team at Facebook

Awni Brian Laurens CrypTen Team at Mark Facebook

Awni Brian Laurens CrypTen Team at Mark Facebook Shobha

Awni Brian Laurens CrypTen Team at Mark Facebook Shobha Shubho

Awni Brian Laurens CrypTen Team at Mark Facebook Shobha Shubho Vini