Highly Reliable Memory-based Physical Unclonable Function Using Spin-Transfer Torque MRAM

Transcription

1 Highly Reliable Memory-based Physical Unclonable Function Using Spin-Transfer Torque MRAM Le Zhang 1, Xuanyao Fong 2, Chip-Hong Chang 1, Zhi Hui Kong 1, Kaushik Roy 2 1 School of EEE, Nanyang Technological University, Singapore 2 School of ECE, Purdue University, USA 1 nd October, 2014

2 Physical Unclonable Function (PUF) Given a stimulus (challenge), produce a response Memory-based PUF Regular structure Available in comp- uter memory system Problems Memory or memory-like array Noise Effects Response Reliability is low (lower in extreme conditions) Overhead of conventional methods, i.e., ECC BL WL Selected cell Challenge=F(memory address) Helper data Fuzzy Extractor Output Key Information reconciliation Privacy amplification

3 Spin-Transfer Torque MRAM based PUF Magnetic Tunneling Junction (MTJ) Free Layer MgO Pinned Layer (a) Parallel Anti-parallel Characteristics Low R High R Normalized MTJ Resistance I C:AP P I C:P AP Applied voltage (V) (b) FL and PL in parallel (P), R MTJ is low FL and PL in anti-parallel (P), R MTJ is high Switching current changes the state of MTJ WL BL SL MTJ (c)

4 Spin-Transfer Torque MRAM based PUF Process variations Cross-section area Oxide layer thickness Process variations of access transistors ( a0tmgo b0 c m 1 2m ( 1) MTJ amtmgo bm d ) m 1 RA e V e Free Layer Threshold MgO voltage, channel Pinned Layer length, etc. Free Layer Cross-section area Measuring the mismatches of two accessed cell resistances generate random bits MgO Oxide-layer thickness Pinned Layer

5 Spin-Transfer Torque MRAM based PUF VDD M13 2. Measure the resistance mismatch by sensing the current difference RCLK WL M11 BIT M7 M9 IreadA M5 RSE M12 BITB M10 M8 M6 IreadB Addr Row Decoder Sense Amplifier Column Decoder PUF cell 1. Initially set MTJ A and MTJ B to the same state Iwrite M3 MTJA RDEN MTJB M4 INIT Internal STT-MRAM Write Driver R (b) H Fuzzy Extractor K BLL SLL M1 (a) M2 BLR SLR (c)

6 Automatic Write-Back (AWB) To improve the reliability, AWB scheme is incorporated Mechanism Enrollment phase: MTJs in a selected cell are initialized to the same state; response bit is generated by sensing the mismatch between them; upon the read completion, response bit is written back to one of the two MTJs Regeneration phase: Response bit is directly regenerated from the cell whose two MTJs have been written to complementary states

7 STT-MRAM based PUF with AWB Circuit block diagram of STT-MRAM PUF incorporated with AWB Data_init Data_wb INIT Sense Amplifier Write Driver PUF architecture 1 0 STT-MRAM Array 0 1 Write-back RDEN SLL RDEN BLL RDEN RDEN SLR RDEN BLR RDEN E: Enrollment R: Regeneration INIT=0 E R MTJ A =P MTJ A =AP or MTJ B =P MTJ B =AP INIT=1 MTJ A =P or MTJ A =P MTJ B =AP MTJ B =AP (State diagram of STT-MRAM with AWB mechanism)

8 Analysis of STT-MRAM based PUF Experiment Set-up HSPICE environment Temperature (T) Supply voltage (V DD ) Process param. (tmgo, pov, Area) pinned MTJ Res model: RMTJ=f(θ,V,T) PMA LLG component MATLAB Data processing for evaluating: Randomness Reliability Material param. (saturation magnetization, damping factor etc.) free MTJ compact model STT-PUF netlist Parameters μ σ/μ Contact area nm 2 5% Oxide layer thickness 1.15 nm 2% MOSFET channel length 45 nm 10% MOSFET threshold voltage V 10% Supply voltage 1 V - Read cycle 10 ns - Evaluations: Randomness and uniqueness Reliability Security analysis Write cycle 40 ns -

9 Uniqueness and Randomness Uniqueness how change of bits in challenge affects the change in response Fractional Hamming distance (H) H Ideally H should be close to 50% m 1 m 2l HD( Ri, Rj) Vbias 0.9 V, 1 V, 1.1 V; l=64, m=1000 m ( m 1) i 1 j i 1

10 Uniqueness and Randomness Randomness how random the generated response bits are Entropy (E) E log[pr( R)]Pr( R) l R {0,1}

11 Reliability Reliability reproducibility of responses to the same challenge Three cases should be considered Case #1: r t and r t+δt are both generated in the enrollment phase; Case #2: r t is generated in the enrollment phase while r t+δt is generated in the regeneration phase; Case #3: r t and r t+δt are both generated in the regeneration phase.

12 Reliability Reliability in Case #1 is mainly affected by thermal noise effects. Reliabilities in Case #2 and #3 may be affected by thermal fluctuations in write-back operations. Distribution of initial angle to the easy axis 2H M k T k s k s 2 Pr( ) exp sin ( ) B H M k T B Bit-Error Rate (BER), i.e., 1-reliability

13 Reliability BER=(BER P +BER AP )/2 Response Surface Modeling Worst-case BER , 10 5 reduction!

14 Efficiency and Security Analysis Comparison to other NVMs, AWB is more applicable to STT-MRAM STT-MRAM has nearly unlimited endurance No disturbance on adjacent cells Resilience against physical attacks small footprint, i.e., 45 nm or even more scaled feature size Resilience against possible side-channel attacks, e.g., electrical-magnetic emanation

15 Conclusion We designed an STT-MRAM based PUF We proposed using an Automatic Write- Back scheme to enhance the PUF reliability The STT-MRAM PUF has desirable qualities: ~50% fractional Hamming distances, entropy of bit per cell, BER in the worst-case More efficient and secure compared to other possible NVM-based PUF implementations

16 Thanks for listening We appreciate your valuable comments