Canada Spends Logo

Explorez notre analyse du budget 2026

Grant: Discovery Grants Program - Individual

The University of British Columbia

$38,000.00

Towards secure acceleration of machine learning models

EF 2024-25

Résumé
With the Cambrian explosion of machine learning applications, dedicated accelerator accelerators are common in both datacentres and edge devices, while bleeding-edge CPUs and GPUs incorporate special tensor instructions. Unfortunately, the same regular execution that makes accelerators efficient allow adversaries to monitor and affect the execution of machine learning (ML) models. Even in datacentre settings, where physical access is difficult to obtain, attackers can remotely reverse-engineer closely guarded models via timing and memory contention attacks. In mobile and IoT devices, physical access is typically much easier, so attackers enjoy even more freedom. They can monitor execution through various side-channels (fine-grained execution timing, memory transaction contents or volume, current draw, thermal signatures, and so on), and can feed carefully crafted inputs to the accelerator to measure responses. This proposal will addresses these challenges in two complementary themes outlined below. The first theme will focus on characterizing attack surfaces of ML accelerators, and developing defences to attacks via hardware side-channels. We will first develop attacks that reverse-engineer the hardware accelerator itself and the models that execute on it; we will then mount attacks that create adversarial inputs. We will rely on a range of side-channels available in ML accelerators, including execution time and memory traffic volume. We will also take advantage of the ability to provide input samples to the accelerators and affect execution, either via sensor inputs or by intercepting off-chip memory traffic. Finally, we will develop countermeasures to attacks that rely on via hardware side-channels, injecting operations to obfuscate side-channels both at the algorithmic level when the workload is mapped to the hardware, and by buffering and reordering side-channel-prone data within the accelerator hardware itself. The second theme will focus on efficient and provably secure ML model execution by developing hardware accelerators for homomorphic encryption (HE) as applied to machine learning workloads. HE offers provable security at the cost of computational efficiency, as even simple computation on encrypted data are expensive. We will focus on taking advantage of weight and activation sparsity during the HE computation; because HE operations take place on large encrypted vectors rather than single values, doing so will require new techniques to reorganize HE data vectors to maximize how much of the available sparsity can be used to skip HE computations. We expect that our research will result in new accelerator architectures that offer resilience to adversarial attacks at the absolute minimum cost in silicon area and energy-efficiency, as well as accelerators that maintain the security guarantees of homomorphic encryption while offering an order of magnitude better performance.
Programme
Discovery Grants Program - Individual
Emplacement
British Columbia, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data

Tel que publié

Application id
RGPIN-2023-05380
Name nom
Lis, Mieszko
Department département
Electrical and Computer Engineering
Organization id
2
Institution établissement
The University of British Columbia
Province en
British Columbia
Province fr
Colombie-Britannique
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2024
Competition year année de concours
2023
Award amount
38000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1510
Committee name en
Electrical and Computer Engineering
Committee name fr
Génie électrique et informatique
Area of application code
513
Area of application group en
Manufacturing processes and products
Area of application group fr
Procédés et produits industriels
Area of application en
Electrical and electronic machinery and equipment (including computer hardware)
Area of application fr
Machinerie et équipement électriques et électroniques (y compris matériel informatique)
Research subject code
2719
Research subject group en
Information technology
Research subject group fr
Technologie de l'information
Research subject en
Computer architecture and design
Research subject fr
Architecture d'ordinateur et conception
Field of research list names en
Not available
Field of research list names fr
Non disponible
Socioeconomic objective list names en
Not available
Socioeconomic objective list names fr
Non disponible
Application title
Towards secure acceleration of machine learning models
Keywords
accelerators; security; encryption; machine learning
Application summary
With the Cambrian explosion of machine learning applications, dedicated accelerator accelerators are common in both datacentres and edge devices, while bleeding-edge CPUs and GPUs incorporate special tensor instructions. Unfortunately, the same regular execution that makes accelerators efficient allow adversaries to monitor and affect the execution of machine learning (ML) models. Even in datacentre settings, where physical access is difficult to obtain, attackers can remotely reverse-engineer closely guarded models via timing and memory contention attacks. In mobile and IoT devices, physical access is typically much easier, so attackers enjoy even more freedom. They can monitor execution through various side-channels (fine-grained execution timing, memory transaction contents or volume, current draw, thermal signatures, and so on), and can feed carefully crafted inputs to the accelerator to measure responses. This proposal will addresses these challenges in two complementary themes outlined below. The first theme will focus on characterizing attack surfaces of ML accelerators, and developing defences to attacks via hardware side-channels. We will first develop attacks that reverse-engineer the hardware accelerator itself and the models that execute on it; we will then mount attacks that create adversarial inputs. We will rely on a range of side-channels available in ML accelerators, including execution time and memory traffic volume. We will also take advantage of the ability to provide input samples to the accelerators and affect execution, either via sensor inputs or by intercepting off-chip memory traffic. Finally, we will develop countermeasures to attacks that rely on via hardware side-channels, injecting operations to obfuscate side-channels both at the algorithmic level when the workload is mapped to the hardware, and by buffering and reordering side-channel-prone data within the accelerator hardware itself. The second theme will focus on efficient and provably secure ML model execution by developing hardware accelerators for homomorphic encryption (HE) as applied to machine learning workloads. HE offers provable security at the cost of computational efficiency, as even simple computation on encrypted data are expensive. We will focus on taking advantage of weight and activation sparsity during the HE computation; because HE operations take place on large encrypted vectors rather than single values, doing so will require new techniques to reorganize HE data vectors to maximize how much of the available sparsity can be used to skip HE computations. We expect that our research will result in new accelerator architectures that offer resilience to adversarial attacks at the absolute minimum cost in silicon area and energy-efficiency, as well as accelerators that maintain the security guarantees of homomorphic encryption while offering an order of magnitude better performance.