A PUF-Integrated Power Amplifier with Harmonic Tunable Matching Network-Controlled Spectral Regrowth Technique for RF Fingerprinting in IoT

loT security has become indispensable and much expected in future wireless networks. A physical layer authentication method called Radio Frequency Fingerprinting (RFF) had been realized through manufacturing process variation [1]. Following which, Q. Zhou et al. proposed the Physical Unclonable Function (PUF)-controlled RFF Power Amplifier (PA) as shown in Fig. 1(a) [2]. This design converted the digital fingerprint from PUF to DAC, and apply it to a PA through tuning the bias voltage to enable random distribution of the Out-ofBand Leakage Power (OOBLP) as unique identity tags. However, such bias control resulted in large In-Band Power (IBP) variation while OOBLP tuning, hence limited its actual applications. Besides, the absolute power of OOBLP is adopted to evaluate the identification error rate, which is channel-dependent and not effective when the signals passed through different paths with different channel losses. To address the above issues, this paper presents a PA integrated with a proposed ultra-low native Bit Error Rate (BER) PUF for RFF application. A harmonic tunable matching network (HTMN) is proposed to achieve a stable IBP when the out-of-band spectrum regrows (Fig. 1(b)). The power difference between OOBLP and IBP ($\triangle$ Power = OOBLP -IBP) is adopted to evaluate the identification error rate to avoid the multipath influence on OOBLP. Compared to the biasing method, the proposed RFF PA with HTMN stabilizes the fundamental output impedance and controls the $2^{\text {nd }}$ harmonic output impedance to realize a very large spectral regrowth tuning and achieve a constant IBP. A 2.4-GHz RFF PA integrated with proposed Static Random-Access Memory (SRAM) PUF is fabricated in $65-\mathrm{nm}$ CMOS process. It has only 0.16 dB maximum IBP variation when the OOBLP changing from -50.8 dBm to -35 dBm with a $2 \mathrm{Mb} / \mathrm{s} \pi / 4-$ DQPSK signal. The identification error rate of 96 data from six chips is only 1.46%.