Skip to content

Running VAEs on mobile and IOT devices using TFLite.

Notifications You must be signed in to change notification settings

UW-CIA/Models_at_edge

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

11 Commits
 
 
 
 

Repository files navigation


contributions welcome Hits License GitHub issues GitHub forks PRs Welcome Open Source Love svg1

Models_at_edge

Running the generative model (VAE) anomaly detector at the edge - IOS/Android/IoT



Part of the following effort: HARNESSING THE POWER OF GENERATIVE MODELS FOR MOBILE CONTINUOUS AND IMPLICIT AUTHENTICATION

Interconnecting the following works:

  • Generative AI Models
  • Continous Authentication
  • Implicit Authentication
  • Outlier Detection

Abstract

Authenticating a user’s identity lies at the heart of securing any information system. A trade off exists currently between user experience and the level of security the system abides by. Using Continuous and Implicit Authentication a user’s identity can be verified without any active participation, hence increasing the level of security, given the continuous verification aspect, as well as the user experience, given its implicit nature. This thesis studies using mobile devices inertial sensors data to identify unique movements and patterns that identify the owner of the device at all times. We implement, and evaluate approaches proposed in related works as well as novel approaches based on a variety of machine learning models, specifically a new kind of Auto Encoder (AE) named Variational Auto Encoder (VAE), relating to the generative models family. We evaluate numerous machine learning models for the anomaly detection or outlier detection case of spotting a malicious user, or an unauthorised entity currently using the smartphone system. We evaluate the results under conditions similar to other works as well as under conditions typically observed in real-world applications. We find that the shallow VAE is the best performer semi-supervised anomaly detector in our evaluations and hence the most suitable for the design proposed. The thesis concludes with recommendations for the enhancement of the system and the research body dedicated to the domain of Continuous and Implicit Authentication for mobile security. Keywords: Machine Learning, Generative Models, Continuous Authentication, Implicit Authentication, Artificial Intelligence

Description

Two different implementations are successfully converted to tflite models that are easily ran at edge thanks to the tflite framework.

REF: https://www.tensorflow.org/lite

Instructions

Releases

No releases published

Packages

No packages published