[Paper Review] Practical Encrypted Computing for IoT Clients

McKenzie van der Hagen, Brandon Lucia|arXiv (Cornell University)|Mar 11, 2021

Cryptography and Data Security54 references4 citations

TL;DR

CHOCO introduces a client-optimized homomorphic encryption system for IoT devices that minimizes communication and client-side computation costs through rotational redundancy and hardware acceleration. With CHOCO-TACO, a custom ASIC accelerator, client encryption/decryption is 1094× faster and 648× more energy-efficient, enabling encrypted DNN inference that uses 37% less energy than local unencrypted computation.

ABSTRACT

Privacy and energy are primary concerns for sensor devices that offload compute to a potentially untrusted edge server or cloud. Homomorphic Encryption (HE) enables offload processing of encrypted data. HE offload processing retains data privacy, but is limited by the need for frequent communication between the client device and the offload server. Existing client-aided encrypted computing systems are optimized for performance on the offload server, failing to sufficiently address client costs, and precluding HE offload for low-resource (e.g., IoT) devices. We introduce Client-aided HE for Opaque Compute Offloading (CHOCO), a client-optimized system for encrypted offload processing. CHOCO introduces rotational redundancy, an algorithmic optimization to minimize computing and communication costs. We design Client-Aided HE for Opaque Compute Offloading Through Accelerated Cryptographic Operations (CHOCO-TACO), a comprehensive architectural accelerator for client-side cryptographic operations that eliminates most of their time and energy costs. Our evaluation shows that CHOCO makes client-aided HE offloading feasible for resource-constrained clients. Compared to existing encrypted computing solutions, CHOCO reduces communication cost by up to 2948x. With hardware support, client-side encryption/decryption is faster by 1094x and uses 648x less energy. In our end-to-end implementation of a large-scale DNN (VGG16), CHOCO uses 37% less energy than local (unencrypted) computation.

Motivation & Objective

• Address the high client-side computational and communication costs in existing client-aided homomorphic encryption (HE) systems, which hinder deployment on resource-constrained IoT devices.
• Overcome the bottleneck of repeated client-side encryption and decryption in client-aided HE, which dominates time and energy costs.
• Design a system that preserves client data privacy without requiring model privacy, enabling efficient offload of complex computations like DNN inference.
• Develop a specialized, low-power hardware accelerator to eliminate the performance penalty of client-side cryptographic operations in HE.
• Demonstrate that encrypted offloading can be more efficient than local inference for IoT devices through end-to-end system evaluation.

Proposed method

• Introduce rotational redundancy, an encrypted permutation algorithm that reduces ciphertext size and communication overhead by enabling smaller, more efficient ciphertexts in client-aided HE.
• Design CHOCO-TACO, a custom ASIC accelerator that implements all core HE cryptographic primitives (e.g., NTT, polynomial multiplication) in hardware to minimize latency and energy.
• Optimize the client-server interaction by offloading linear operations (e.g., convolutions) to the server using HE, while performing non-linear operations (e.g., ReLU) in plaintext on the client.
• Use a hybrid architecture where the client handles only non-linear operations, avoiding the need for MPC and thus eliminating its high client-side cost.
• Implement a full-stack system integrating software optimizations and hardware acceleration, targeting low-power IoT platforms with Bluetooth communication.
• Evaluate the system end-to-end using VGG16 inference, comparing performance, energy, and communication costs against software-only HE, MPC-based systems, and local TFLite inference.

Experimental results

Research questions

• RQ1 Can client-aided HE be made practical for resource-constrained IoT devices by reducing client-side communication and computation costs?
• RQ2 How effective is rotational redundancy in minimizing ciphertext size and communication overhead in client-aided HE?
• RQ3 To what extent can hardware acceleration of HE primitives reduce the time and energy cost of client-side encryption and decryption?
• RQ4 Can encrypted offloading via CHOCO outperform local inference in terms of energy efficiency and performance on IoT devices?
• RQ5 Is it feasible to achieve privacy-preserving DNN inference on IoT clients without requiring model privacy or MPC overhead?

Key findings

• CHOCO reduces client communication cost by up to 2948× compared to existing HE-MPC protocols, primarily through smaller ciphertexts enabled by rotational redundancy.
• With CHOCO-TACO hardware acceleration, client-side encryption/decryption is 1094× faster and consumes 648× less energy than in software-only HE systems.
• In end-to-end VGG16 inference, CHOCO uses 37% less energy than local unencrypted inference using TFLite, demonstrating energy efficiency gains.
• CHOCO-TACO achieves a 417× speedup and 603× energy reduction for single encryption operations, translating to a 121× average speedup in client-side DNN inference.
• The system enables encrypted offloading to be competitive with local inference in performance and energy efficiency, even on low-power IoT devices.
• CHOCO-TACO's comprehensive hardware support for all HE primitives is critical, as NTT acceleration alone is insufficient to overcome client-side bottlenecks.