Aegis: A Single-Chip Secure Processor

Abstract

Trust in remote interaction is a fundamental challenge in distributed computing environments. To obtain a remote party's trust, computing systems must be able to guarantee the privacy of intellectual property and the integrity of program execution. Unfortunately, traditional platforms cannot provide such guarantees under physical threats that exist in distributed environments.The AEGIS secure processor enables a physically secure computing platform to be built with a main processor as the only trusted hardware component. AEGIS empowers a remote party to authenticate the platform and guarantees secure execution even under physical threats. To realize the security features of AEGIS with only a single chip, this thesis presents a secure processor architecture along with its enabling security mechanisms. The architecture suggests a technique called suspended secure processing to allow a secure part of an application to be protected separately from the rest. Physical random functions provide a cheap and secure way of generating a unique secret key on each processor, which enables a remote party to authenticate the processor chip. Memory encryption and integrity verification mechanisms guarantee the privacy and the integrity of off-chip memory content, respectively.A fully-functional RTL implementation and simulation studies demonstrate that the overheads associated with this single-chip approach is reasonable. The security components in AEGIS consumes about 230K logic gates. AEGIS, with its off-chip protection mechanisms, is slower than traditional processors by 26% on average for large applications and by a few percent for embedded applications. This thesis also shows that using AEGIS requires only minor modifications to traditional operating systems and compilers.

Lightweight and Secure PUF Key Storage Using Limits of Machine Learning

“…Lightweight ECC has an implementation complexity that is estimated to be 75% smaller than the two-stage scheme published in [21] (secure based on i.i.d. PUF output assumption) and an estimated 98% smaller than the single-stage scheme published in [18] (secure based on Dodis' framework). The results are summarized in Table 1 below.…”

“…It achieves a 75% reduction in ECC complexity compared to [21] and achieves a PUF complexity of 5 (using ten 64-sum PUFs, see Figure 1 for a description of k-sum PUFs) to as little as 1 (using two 64-sum PUFs); this is a 4x to 20x improvement. The PUF complexity reduction derives from a machine-learningbased security argument that each additional syndrome bit does not require a linear increase in the number of PUF elements (e.g., disjoint oscillator pair [18] [21] or a memory cell [1] [12] [7] [17] [8]) but instead relies on assumptions on what cannot be learned about a challengeable physical system.…”

Reliable and efficient PUF-based key generation using pattern matching

“…The work employed 2D Hamming codes for error correction. Later, a more realistic use of Bose-Chaudhuri-Hochquenghen (BCH) codes for error correction on PUF responses was proposed [9], [20]; however, the implementation cost and hardware overhead of this code is very high.…”

“…In conventional usage of a PUF as a key generator, only a fixed number of secret bits need to be generated from the PUF. These bits can be used as symmetric key bits or used as a random seed to generate a public/private key pair in a secure processor [9]. However, in order for the PUF outputs to be usable in cryptographic applications, the noisy bits need to be error corrected, with the aid of helper bits, commonly referred to as a syndrome.…”

Autonomous Physical Secret Functions and Clone-Resistant Identification

“…This fact leads to the idea of accumulatively changing/evolving the hash function as a part of the device personality and keep tracing its evolution generations. The same identification technique is used for PUF-based identification [1][2][3] with the difference that the secret seed key and the hash mapping are unchangeable as they are physically inherent unknown intrinsic mappings. Identification using PUFs is actually the ultimate solution if the inherent properties are consistently reproducible (i.e.…”