DVFS Frequently Leaks Secrets: Hertzbleed Attacks

Maatla Sefawe and Ayush Vashi

1

Background

2

DVFS

• Modern CPUs dynamically adjust frequency to reduce power consumption.
• When the CPU load is low, frequency is decreased to save power.
• Can be used as a throttling mechanism when CPU load is high.

3

4

Hamming Weight

• Hamming weight measures the # of ones in a string of bits
• E.g. 001110 has a hamming weight of 3

Hamming Distance

• Hamming distance measures the number of bits that flip when changing a value
• E.g. The hamming distance between 0100 and 1101 is 2

Hertzbleed

• A class of timing attacks that leverage the discovery that, during high CPU loads, DVFS-induced frequency adjustments depend on the CPU power consumption the data being computed

​

• This is because frequency adjustments rely on data-dependent CPU power consumption

​

• Data-dependent frequency changes translate directly to remotely observable execution time differences

5

• Traditional timing attacks leak “digital features” (e.g, instruction count, branches taken)
• Hertzbleed leaks “analog features” (e.g, temperature, power)

6

Impact

• This finding fundamentally challenges constant-time programming principles designed to mitigate traditional timing and power side channels attacks

7

Expanding the Threat Scope

• Initially only demonstrated on a specific cryptographic primitive, SIKE

8

Expanding the Threat Scope

9

10

Elliptic Curve Digital Signature Algortihm(ECDSA)

• Leaked number of leading 0 bits in the ephemeral nonce
• Operations on low HW data cause lower CPU power consumption, resulting in a higher CPU frequency
• Not very practical due to small timing differences and noise

Classic McEliece

• A decoding failure produces an array with far fewer non-zero entries compared to a successful attempt
• Bitslicing in the re-encryption check amplifies this into a significant low HW/HD data difference, leading to higher CPU frequency
• Demonstrated full plaintext recovery

Beyond Cryptography and CPU Core Data: Leaking Web Browser Secrets from the iGPU

11

iGPU-CPU Frequency Leakage Channel

• CPU frequency adjustments are affected by the iGPU power consumption
• iGPU power consumption depends on the data values being computed within the iGPU

​

12

13

CPU Frequency is iGPU workload-dependent

CPU Frequency is iGPU data-dependent

Pixel Stealing

• Web browsers manage secrets on behalf of users
• Some secrets influence visible content on the page
• “Pixel stealing is a cross-origin leak where the secret revealed is the visual content of a target site page.”

14

SVG Filter Stack Framework

• A series of graphical computations that can be applied to elements of a web page to create visual effects
• Google Chrome runs these by default
• The goal is to find a stack which when rendering on an iframe filled with an attacker-selected cross-origin pixel, creates iGPU power differences

15

Attacker Setup

• Target a single black or white pixel in a cross-origin iframe
• Use CSS properties to isolate and expand this single pixel into a large uniformly colored iframe

16

Filter Stack

1. Generate a large random image of random colors
2. Apply an arithmetic composition between the target iframe and the random image

17

Amplification

• Using the composed image as a source, append GPU heavy filter feGaussianBlur
• This works as an amplifier of the HW/HD difference in the composed images
• The iGPU power consumption should be very distinct depending on the color of the target pixel

18

Measuring CPU Frequency

• The accumulated counter represents the relative CPU frequency between the two clock edges

19

Experimental Results

20

21

• Applied their FilterStack repeatedly on the iFrame targeting either a white or black pixel in an infinite loop
• Simultaneously, stressed the CPU by running a JavaScript worker thread per core

Proof of Concept

22

23

24

Discussion

25

26

What are the possible software mitigations against this attack?

Question 1

Browser Side Mitigations

• Restricting or removing the application of graphical transforms on cross-origin information
• Disabling cross-origin iframe access to sensitive data(like cookies)
• Re-implementing SVG filters to be resilient against this type of power leakage

27

28

What are the possible hardware mitigations against this attack?

Question 2

29

What are the broader security implications of cross component leakage (iGPU-CPU leakage) on modern processors?

Question 3

Thank you!

30