FP-Rowhammer: DRAM-Based Device Fingerprinting

TL;DR

FP-Rowhammer is the first Rowhammer fingerprinting approach to extract unique and stable fingerprints efficiently and at scale and is the first Rowhammer fingerprinting approach to extract unique and stable fingerprints efficiently and at scale.

Abstract

Device fingerprinting leverages attributes that capture heterogeneity in hardware and software configurations to extract unique and stable fingerprints. Fingerprinting countermeasures attempt to either present a uniform fingerprint across different devices through normalization or present different fingerprints for the same device each time through obfuscation. We present FP-Rowhammer, a Rowhammer-based device fingerprinting approach that can build unique and stable fingerprints even across devices with normalized or obfuscated hardware and software configurations. To this end, FP-Rowhammer leverages the DRAM manufacturing process variation that gives rise to unique distributions of Rowhammer-induced bit flips across different DRAM modules. Our evaluation on a test bed of 98 DRAM modules shows that FP-Rowhammer achieves 99.91% fingerprinting accuracy. FP-Rowhammer's fingerprints are also stable, with no degradation in fingerprinting accuracy over a period of ten days. We also demonstrate that FP-Rowhammer is efficient, taking less than five seconds to extract a fingerprint. FP-Rowhammer is the first Rowhammer fingerprinting approach that is able to extract unique and stable fingerprints efficiently and at scale.

Figures (13)

• Figure 2: Plots showing the variation in the number of bits of entropy that can be obtained to represent different regions of memory across a trillion DIMMs with varying number of bit flips.
• Figure 3: Set A shows the addresses of bits that flipped when hammering a particular chunk of a DIMM. Set B shows the addresses of bits that flipped when restoring data to the chunk and hammering it again. The two sets are disjoint demonstrating the non-deterministic behavior of bit flips.
• Figure 4: Set U shows the set of all addresses that flipped across 8 attempts to restore the data and hammering the same chunk from Figure \ref{['fig:disjoint']}. The boldfaced addresses are addresses that flipped more than once across the 8 attempts.
• Figure 5: Visualization of the relative persistence of bit flips within given 2 MB chunks of memory across multiple DIMMs.
• Figure 6: Visualization of a hammering sweep performed within one bank of a contiguous 2 MB chunk. The central rectangular blocks in the visualization represent rows within the chunk. We map primary aggressors in the discovered patterns to rows within the chunk and secondary aggressors to random rows within the same bank. The hammering sweep involves sequentially hammering all pairs of double-sided aggressors as primary aggressors within the bank of the chunk and scanning the other rows for bit flips.