A Sensor-Assisted Self-Authentication for Hardware Trojan Detection

Min Li*, Azadeh Davoodi*,

Mohammad Tehranipoor**

*University of Wisconsin-Madison

**University of Connecticut

WISCAD Electronic Design Automation Lab http://wiscad.ece.wisc.edu

2

Challenges of Hardware Trojan Detection• Challenges:

– lack of observability and controllability after fabrication– complexity

• due to existence of billions of nano-scale components• due to high volume of soft and hard integrated IP cores

– overhead associated with physical inspection of nanometer feature sizes for reverse engineering

• could be intrusive

– difficulty to activate a Trojan– increasing fabrication and environmental variations with

technology scaling

3

Fundamental Challenge• Trojan-free or Golden IC (GIC)

– required in any (generic) IC authentication process• create a reference fingerprint from the transient behavior of GIC and

compare with fingerprint obtained from target IC

– existence and identification of GIC cannot be guaranteed• if inserted in GDSII file, or if the foundry alters the mask to insert a

Trojan, GIC will not exist• if an IC passes a rigorous test, in theory one cannot conclude that it is

a GIC

4

Contributions• A framework to use custom-designed on-chip detection

sensors to alleviate the need on a golden IC by providing a self-authentication

• On-chip “detection sensor”– a compact (small area) representation of a design

• can be designed by searching for common “features” in a design

– shares common sources of uncertainty with the design due to realization on the same chip

• e.g., process and environmental variations

• Assumptions – Trojan may infect the design paths, the detection sensor, or both

– the detection sensor is obfuscated within the design’s layout• i.e., an adversary will not be able to distinguish it

5

Proposed Framework

Design and integration of custom-generated detection

sensorscapturing within-die variability

Design stage

On-chip delay fingerprint of

detection sensors

On-chip delay fingerprint of

arbitrary design paths

PASS?

Alert Trojan

Offline analysis of fingerprint correlation

NO

Post-silicon self-authentication process

6

Design stage

Post-Silicon

detection sensor: a compact

representative of the design

measured on-chip delays of design

paths and of detection sensors

analyze correlation

-- Finds most frequent “layout features” which are design-dependent and technology-sensitive

-- Main Idea: Addition of Trojan disturbs the expected delay correlation between the design and the detection sensor

detect Trojan

7

Design of the Detection Sensor• Steps

– logic design via netlist analysis1. sequence matching

2. feature discovery

– physical design of sensors• layout integration

• delay measurement

8

Design of A Detection Sensor• An optimization framework for finding frequent

sequences in a netlist*– modeled using a graph representation of the design’s netlist– break (all or some portions) of the graph into collection of “similar”

sequences– similar sequences grouped together and represented by one

sensor– given a budget for total area used by the detection sensors, the

goal is to maximize coverage of graph with formed sequences

*Li and Davoodi, “Custom On-Chip Sensors for Post-Silicon Failing Path Isolation in the Presence of Process Variations”, Technical Report, 2011

9

Design of the Detection Sensor• Constraints:

1. sequence constraints• a sequence is made of

consecutive edges

2. similarity constraints• “similar” sequences are

mapped to the same sensor• similarity defined based on

delay correlation in the presence of uncertainties such as process variations

3. area constraint• summation of the areas of

detection sensors are bounded

original netlist extended graph after modification

simplified version for illustration

10

Variation-Aware Delay Modeling

[Agarwal et al, ASPDAC’03]

2 , 2

2 , 5

2 , 6

2 , 3

2 , 4

2 , 7

2 , 8

2 , 9

2 , 1 0

2 , 1 1

2 , 1 2

2 , 1 3

2 , 1 4

2 , 1 5

2 , 1 6

1 , 1

1 , 2

1 , 3

1 , 4

2 , 1

0 , 1

2,1 1,1 0,1a aL L L L r 2,4 1,1 0,1b bL L L L r 2,15 1,4 0,1c cL L L L r

a

c ba

a a La ad s L

b b Lb bd s L

c c Lc cd s L

11

Design of A Detection Sensor• Benefits of the formulation

– flexible definition of similarity between sequences mapped to the same sensor,

• for example similar sequences could be:– structually identical (e.g., same sequence of logic gates)– have the same timing distribution– highly correlated in their timing characteristic– in general can define similarity with respect to sensitivity to technology

parameters

– flexible objective• if edge weights are equal, objective is maximizing netlist coverage• can modify to also ensure spatial coverage from different regions of

the chip

12

Design

Post-Silicon

detection sensor: a compact

representative of the design

measured on-chip delays of design

paths and of detection sensors

analyze correlation

detect Trojan

-- e.g., BIST technology

-- First check BIST is healthy (e.g., by verifying delays of embedded ring oscillators)

13

• Examples– Path-RO [Tehranipoor et al ICCAD08]

• requires inserting measurement circuitry at the pre-silicon stage along the desired representative paths

– Shrinking clock signal [Abraham et al GLSVLSI 2010]

On-Chip Path Delay Measurement

14

Design

Post-Silicon

detection sensor: a compact

representative of the design

measured on-chip delays of design

paths and of detection sensors

analyze correlation

detect Trojan

-- Detection scenarios:• Trojan may be added

to the design, the detection sensor, or both

-- The timing correlation between the design and its detection sensors will be different in the presence of a Trojan

15

Detection Scenarios1. Trojan inserted in the design paths

– actual delay range: obtained from direct path delay measurement considering measurement error

– predicted delay range: computed using actual sensor delay and predicting the remainder of the path using worst/base-case values

Trojan-infected path used for actual delay range

path used for predicted delay range

sensor matched

case-based estimate

sensor

16

Detection Scenarios1. Trojan inserted in the design paths

– underestimation of path delays that are Trojan infected– detects Trojan if predicted delay range does not overlap with the

measured range

sensor

17

Detection Scenarios2. Trojan inserted in the detection sensor

– actual delay range: obtained from direct path delay measurement considering measurement error correct range

– predicted delay range: computed using Trojan-infected sensor delay and predicting the remainder of the path using worst/base-case values

path used for actual delay range

path used for predicted delay range

sensor matched

case-based estimate

sensor

18

Detection Scenarios2. Trojan inserted in the detection sensor

– overestimation of the predicted range of the design paths– correctly detects existence of Trojan if the two ranges don’t

overlap– can identify that detection sensor is infected

sensor

19

Detection Scenarios3. Trojan inserted simultaneously in the design and

detection sensor– actual and predicted ranges both erroneous– depending on how the Trojan impact each one, different cases

can happen

sensor

path used for actual delay range

path used for predicted delay range

20

Detection Scenarios3. Trojan inserted simultaneously in the design and

detection sensor– can only predict that Trojan exists if the predicted and measured

ranges do not overlap, otherwise it doesn’t generate any output

sensor

21

Simulation Setup• Randomly selected a subset of critical design paths from

the ISCAS89 suite• For each considered path

– inserted a Trojan at a random location on the path– sensor area budget is 15%– repeated many times for varying Trojan delays

• 3 to 10% of the delay of the longest path in the circuit (30 different values uniformly selected from the range)

– assumed on-chip measurement error of 3% for measuring the delays of the infected paths

– variation modeling• assumed process variations in channel length and threshold voltage

of transistors according to variation setting for 45nm technology and a 5-level spatial correlation model

• considered 10K scenarios of variations

22

Trojan Inserted in Design

sensor and path information DR (Trojan in design)

Bench |P| %Asensor MR W/O sensors Sensor-assisted

s1423 92 13 1 0.58 0.68

s1488 16 12 1 0.60 0.67

s1494 16 12 1 0.48 0.50

s5378 201 12 0.84 0.62 0.66

s9234 169 10 1 0.69 0.71

s13207 331 12 1 0.67 0.73

s15850 136 15 0.97 0.75 0.85

s35932 1765 13 0.96 0.70 0.73

s38417 1587 12 0.99 0.72 0.77

s38584 1112 14 1 0.64 0.80

MR: fraction of the paths which are matched with at least one sensorDR: detection ratio

23

Trojan Insertion in Detection Sensorsensor and path information DR (Trojan in design)

Bench |P| %Asensor MR W/O sensors Sensor-assisted

s1423 92 13 1 0.03 0.67

s1488 16 12 1 0.04 0.69

s1494 16 12 1 0.04 0.60

s5378 201 12 0.84 0.03 0.58

s9234 169 10 1 0.01 0.75

s13207 331 12 1 0.01 0.77

s15850 136 15 0.97 0.01 0.70

s35932 1765 13 0.96 0.01 0.93

s38417 1587 12 0.99 0.02 0.78

s38584 1112 14 1 0.00 0.97

MR: fraction of the paths which are matched with at least one sensorDR: detection ratio

24

Trojan Detection Rate in s13207

% Trojan delay/delay of longest path

detection rate

25

Conclusions• Benefits of the detection framework

– alleviates the need on a GIC– does not output a wrong answer– detection at a finer granularity– faster detection of Trojan– captures both layout-dependency and technology-dependency

• Limitations– spatial correlation modeling– path delay measurement accuracy– layout integration