1 of 29

CS-773 Paper Presentation��Maya: Falsifying Power Side Channels with Dynamic Control

�

Rishi Agarwal�ArchMages(#0)

rishiagarwal@cse.iitb.ac.in

1

Maya: Falsifying Power Sidechannels with Dynamic Control (arxiv.org)

2 of 29

Motivation

Keystrokes
Password lengths
Location
Browser and camera activity
Application activity
Bits of encryption keys

2

Keystroke

Password length

Location

Browser activity

Camera activity

Application activity

Bits of encryption key

3 of 29

Motivation

Detecting application through power measurement

3

Timesteps

Avg. Power (W)

4 of 29

Motivation

Reshape power in application transparent manner

4

Avg. Power (W)

Timesteps

5 of 29

Prior Work

Require new hardware
Trusted execution environment ineffective against physical side channels
Simple distortion by adding noise is ineffective

5

6 of 29

Maya - 10K feet view

6

Mask generator

Controller

Computer system

Target power (R)

Total power (P)

Inputs

DVFS
Idle thread
Balloon thread

7 of 29

Maya - Controller

7

Controller

E(t) = P(t)-R(t)

S(t+1) = A x S(t) + B x E(t)

A(t) = C x S(t) + D x E(t)

R(t)

P(t)

E(t) := error(t), S(t) := State(t), A(t) := action(t)

A(t)

8 of 29

Matrix multiplication in controller?

Conventional environment

Coarse granular (~20ms)
Easy to mount, widely used

Specialized environment

Fine granular (~10ns)
Expensive and difficult

Matrix based controller responds in 5-10 us
Cannot be used for Specialized environment
Alternatives?

Use a table of pre-computed values

8

9 of 29

How good is the controller?

9

High fidelity power shaping with control theory

Avg. Power (W)

Timesteps

Target power R(t)

Measured power P(t)

10 of 29

Maya - Mask Generator

Mean and variance must change over time domain
Frequency domain curve must spread over a range of frequencies
Peaks must exist in the frequency domain curve

Philosophy?

Applications show these properties
If mask does not exhibit these, controller will not be able to hide them

10

An effective mask must protect information leaked in both time domain and the frequency domain (i.e., after obtaining its FFT) because attackers can analyze signals in either domain.
You need an effective mask generator to make the controller’s life easy. i.e. you cannot assume that the controller can reshape the output power to any desired target power perfectly. If the target power is very difficult to be achieved there will again be information leakage. So, how to generate target power that is effective?
The authors propose the following three requirements of an effective mask generator

Its mean and variance must change over the time domain, to portray phase behavior
The second property is that the mean and variance changes must have various rates – from smooth to abrupt. This property will cause the resulting frequency domain curve to spread over a range of frequencies
The final property is that the target signal must have repeating patterns at various rates. This is to create various peaks in the frequency domain curve (Figure 4(e) bottom). Such peaks are common in applications, as they represent the effects of loops.

11 of 29

Constant Mask

Time: constant mean and 0 variance
Frequency: no spread/ no peaks

11

Time domain

Frequency domain

Power (W)

Time (s)

Freq. (Hz)

12 of 29

Gaussian Sinusoidal Mask

Time: varying mean & variance
Frequency: spread and peaks

12

Time domain

Frequency domain

Power (W)

Time (s)

Freq. (Hz)

13 of 29

Maya - Mask Generator

13

An effective mask must protect information leaked in both time domain and the frequency domain (i.e., after obtaining its FFT) because attackers can analyze signals in either domain.
You need an effective mask generator to make the controller’s life easy. i.e. you cannot assume that the controller can reshape the output power to any desired target power perfectly. If the target power is very difficult to be achieved there will again be information leakage. So, how to generate target power that is effective?
The authors propose the following three requirements of an effective mask generator

Its mean and variance must change over the time domain, to portray phase behavior
The second property is that the mean and variance changes must have various rates – from smooth to abrupt. This property will cause the resulting frequency domain curve to spread over a range of frequencies
The final property is that the target signal must have repeating patterns at various rates. This is to create various peaks in the frequency domain curve (Figure 4(e) bottom). Such peaks are common in applications, as they represent the effects of loops.

14 of 29

Assumptions & Strengths

Assumptions

Conventional environment
Attackers know the algorithm
Hardware or privileged software is uncompromised

OS scheduler
DVFS interfaces

Strengths

Reshaping power protects from temp./EM leakage
Software implementation possible

14

15 of 29

Maya - Summary

Matrix based controller

S(t+1) = A x S(t) + B x E(t)
A(t) = C x S(t) + D x E(t)

Mask generator - gaussian sinusoidal

15

16 of 29

Experimental setup

2 machines - consumer class, server class
Three possible threads

Parallel benchmark
Idle thread
Balloon thread - floating point array ops

Power measured through RAPL
Controller, mask generator, balloon program - privileged

16

17 of 29

Actuating the inputs

DVFS: using cpufreq
Idle thread: using Intel’s Power Clamp interface
Balloon program: using shm file

17

18 of 29

Implementation - controller

Controller

Dynamic model of system
Three parameters

Input weights (1,1,1)
Uncertainty guardband (40%)
Output deviation bounds (10%)

Matrices A,B,C,D generated

18

19 of 29

Dynamic model of system

System identification
y(t) = ∑_{1 … m}a_iy(t-i) + ∑_{1 … n}b_i u(t-i+1)

19

20 of 29

Implementation - mask generator

R(t) = offset + Amp x sin (2𝝿 t / Freq) + N(μ, σ)
Constraints

Thermal design power
Freq < Half of power sampling freq.

N_hold samples

20

21 of 29

Evaluation methodology

Train ML models - MLP
2 attacks

Data w/o Maya
Data with Maya (adaptive attack)

Classification accuracy

21

22 of 29

Accuracy - Data w/o Maya

22

True label

Predicted label

Acc. drops from 92% to 12% (constant), 13%(gaussian-sinusoidal)

Constant mask

Gaussian-sinusoidal mask

23 of 29

Accuracy - Data w/ Maya

23

True label

Predicted label

Acc. increases to 67% for constant mask
Remains as low as 20% for gaussian-sinusoidal

Constant mask

Gaussian-sinusoidal mask

24 of 29

Visualization of power trace

24

Maya reshapes power in application transparent manner
Constant mask less effective than gaussian-sinusoidal

Avg. Power (W)

Timesteps

Constant mask

Gaussian-sinusoidal mask

No Maya

25 of 29

Performance - power

25

Avg. power consumed is 35% (noisy), 41% (constant), 29% (gaussian-sinusoidal) lower than normal

Normalized Power

Applications

26 of 29

Performance - execution time

26

Avg. slowdown of 63% (noisy), 100% (constant), 50% (gaussian-sniusoidal)

Normalized Duration

Applications

27 of 29

Summary

Maya - i) Controller ii) Mask generator

Acc. drops to 13% (non-adaptive), 20% (adaptive) as compared to 92%
Avg. power consumed 29% lower
Avg. execution time 50% more

Can be implemented on software
Power reshaping protects from temp./EM

27

28 of 29

Discussion

Effect of context switch on controller?
More experiments using advanced ML - transformers?
How well do matrices generalize to new applications?
A short note on control theory can be helpful

Not clear how addition of error(t) with State(t) is done

28

29 of 29

��Thank you for listening!

Any Questions?

�

29