A Low-Power Biomimetic Crypto Engine for All-In-One IoT based on Programmable and Multifunctional MoS2 FETs


 In the emerging era of internet of things (IoT), ubiquitous sensors continuously collect, consume, store, and communicate an astonishing volume of information, which are becoming increasingly vulnerable to theft and misuse. Modern software crypto systems are powerful but require extensive computational infrastructure for implementing ciphering algorithms making it difficult to be adopted by IoT edge sensors that operate with limited hardware resources and at miniscule energy budgets. Here we propose, and experimentally demonstrate a low-power, biomimetic, crypto system integrated with IoT edge sensor based on an array of atomically thin, multifunctional, and programmable MoS2 field effect transistors (FETs). We show that the information received by a MoS2 photodetector and encrypted by a population of MoS2 based reconfigurable artificial neural encoders is secure from an eavesdropper with finite resources. We also show that our all-in-one IoT platform consumes miniscule energy in the range of tens to hundreds of pico Joules, has a small hardware footprint, and combines sensing, non-volatile storage, and security, for the first time.


Introduction
Information security is key for sustainable growth and development of any modern society that thrives on global connectivity in this new era of Internet of Things (IoT). Today, information is collected, stored, and communicated continuously by IoT sensors and edge devices that are found ubiquitously in our homes, workplaces, industrial manufacturing plants, transportation, health sectors, agricultural fields, and so on and so forth. However, there is an escalating threat of information loss, misuse, and manipulation owing to the involvement of untrustworthy parities [1].
While the state-of-the-art crypto systems offer powerful security solutions based on complex ciphering algorithms [2][3][4] that can be implemented using hardware accelerators [5], IoT edge devices have many restrictions in terms of computational capabilities due to limited hardware and energy resources. Furthermore, low-cost design needs, large-scale deployments, and heterogeneous nature of the IoT sensors limit direct adoption of traditional security solutions, including the widely used public key scheme. Due to inadequate security, IoT devices used in smart-cars, and smart-homes have shown tremendous vulnerability in the recent times [6].
Therefore, wider adoption of IoT technology can be greatly hindered if cryptosystems and security protocols which require less computational resources are not developed and integrated with the IoT edge sensor in a timely manner.
Here, we exploit a new paradigm, namely, in-memory biomimetic computing to offer an integrated sensing, storage, and security solution for IoT edge devices with minimal hardware investments and at frugal energy expenditure. Our demonstration is based on atomically thin and multifunctional MoS2 field effect transistors (FETs) with a programmable gate stack that can be used as sensor, i.e. photodetector (PD), as well as various components of the proposed cryptographic engine or artificial neural encoder. The encryption is done by a finite population ( ) of encoders with reconfigurable encoding threshold ( ) using a zero mean white Gaussian noise (WGN) of finite standard deviation ( ). The decryption requires an optimum number of voting mandate ( ) that is determined by , , and without the knowledge of which an eavesdropper requires an astronomical number of brute force trials (BFTs) for deciphering the information. In fact, the information remains concealed even if the eavesdropper has access to a trained artificial neural network (ANN). Note that our inspiration is derived from the organization of peripheral and  which are used as photodetectors (PDs) for sensing, and WGNAs and ANs for encryption. Fig. 1c shows an example experimental demonstration of sensing and ciphering. Information, for example, an 8×8 pixelated image of the letter 'N' obtained by illuminating a blue light emitting diode (LED) is presented to the IoT sensor, i.e. MoS2 PD. The photocurrent (IPH) in response is superimposed  with zero mean WGN of desirable standard deviation and transduced to subthreshold presynaptic voltage (VPSV) using MoS2 WGNAs and presented to MoS2 ANs with pre-programmed threshold voltages. The information is revealed by a decoder through a voting process if the encoding knowledge is accessible. see Extended Data 2 for the description of the experimental setup and measurement procedures.
The use of MoS2 for our all-in-one IoT platform is motivated by recent studies demonstrating various low-power sensors based on MoS2 that can benefit the IoT technology [7][8][9][10][11][12][13]. In addition, MoS2 offers compatibility with flexible [14] and printable technologies [15] and shows promise for neuromorphic and biomimetic applications [16][17][18]. MoS2 used in this study was grown epitaxially on a sapphire substrate using metal organic chemical vapor deposition (MOCVD) technique at 1000 0 C and subsequently, transferred from the growth substrate to the device fabrication substrate using the PMMA-assisted wet transfer process [19]. The large area MOCVD growth allows for the fabrication of low-power and programmable monolayer MoS2 FET arrays that can be used for sensing, storage, and ciphering information. See Methods section for further details on the synthesis, film transfer, and fabrication of MoS2 FETs. Fig. 2a Pt/TiN/p ++ -Si. As we will discuss next, this gate stack allows realization of analog, non-volatile, and programmable memory states in MoS2 FETs, which is the key towards the realization of crypto engine for IoT security. In addition, the use of thin and high-k gate oxide such as Al2O3 compared to conventional 300 nm of SiO2 facilitates better electrostatic control of the MoS2 channel and allows operation below 5 V, which is critical for achieving low-power IoT platform. As seen in Fig. 2c MoS2 FET is a unipolar, and n-type thresholding device with = 0.8 V, extracted for = 10 pA/µm, which is 10 times higher than the average noise floor, i.e. 1 pA/µm. In other words the device is considered to be ON if ≥ 10 pA/µm. The device also exhibits excellent ON/OFF current ratio of ~10 7 and subthreshold slope (SS) of less than 225 mV/decade. The electron field effect mobility ( ) value extracted from the peak transconductance was found to be ~10 cm 2 /V-s. Fig. 2d  Note that while our mobility and ON current values are on par with the state-of-the-art literature on large area grown MoS2, these do not play a significant role in our proposed IoT platform as we will exploit subthreshold device operation to achieve energy efficiency.
Next, we demonstrate the capability of programming our monolayer MoS2 FETs in any desirable conductance state with non-volatile retention characteristics. The results are shown in Fig. 2e-h.
When "Write" programming pulses of different amplitudes, , are applied to the back-gate electrode, each for a total duration of = 1 s, the transfer characteristics of the device shifts towards the right as illustrated in Fig. 2e. During programming, the source and drain terminals were grounded. Fig. 2f shows the extracted iso-current (~ 10 pA) threshold voltages, , corresponding to each state measured multiple times, post-programming, to ensure non-volatile retention. Similar observations are made when "Write" programming pulses of same amplitude, = 10 V, but different tP, are applied to the back-gate electrode as shown in Fig. 2g. Fig. 2h shows the corresponding non-volatile shift in . The shift in can be attributed to our backgate stack that closely resembles floating gate (FG) configuration used in non-volatile flash memory [20]. See Extended Data 3 explaining the memory operation using energy band diagrams.
In short, the p ++ -Si/TiN/Pt interface in the stack is characterized by a Schottky barrier (SB), whereas, the gate dielectric, i.e. 50 nm Al2O3, acts as an oxide barrier (OB). The OB is much wider and taller compared to the SB. When a large positive back-gate voltage, = , i.e. "Write" pulse is applied to the control gate (CG), i.e. p ++ -Si, carriers tunnel from the p ++ -Si into the Pt/TiN floating gate (FG) and remains trapped even when the is released. These negative fixed charges on the FG screen the electric field from CG and thereby makes the more positive. The total amount of charge injected into the FG, and hence shift in of the MoS2 FET can be controlled by the amplitude, and duration, of the "Write" programming pulse as shown in Fig. 2e and 2g, respectively. Note that once programmed, the devices continue to remain in the programmed state as evident from the retention measurements displayed in Fig. 2f and 2h. This is critical for nonvolatile memory operation. Furthermore, the device can be programmed in any desired state indicative of analog memory operation, which we will exploit later for the realization of look-up- Next, we demonstrate the photosensing capability of monolayer MoS2 FET. Fig. 2i shows the transfer characteristics of a representative MoS2 FET in dark and under the illumination of a blue LED, which is placed at ~ 1 cm distance operating at its maximum rated brightness (5 V). Clearly, the device shows reasonable photoresponse and hence can be used as a photodetector (PD). Note that unlike most studies that use LASER excitation to evaluate the photoresponse of MoS2 FETs, we have used LED as the optical source to resemble more realistic lighting ambience where most IoT sensors will be deployed. The phototransduction mechanism in MoS2 PD is extensively studied in the literature including our previous reports and can be ascribed to a combination of photocarrier generation in the MoS2 channel as well as photogating effect arising due to charge trapping/detrapping at the MoS2/gate-dielectric interface [21]. demonstrate that the MoS2 PD is able to accurately transcribe the optical information into electrical response. Note that the MoS2 PD was biased in the subthreshold regime to enable exponential reduction in the dark current (~ 1 pA) and thereby making = under illumination. This also allows ultra-low-power photodetection with energy expenditure in the range ~25-30 fJ/pixel, averaged over all pixels.  integrates well with our all-in-one IoT platform. Fig. 3d shows the transduction of the photocurrent map for the letter 'N' to for different noise standard deviation ( ). Note that the resistive network used by the CA allows linear transformation of the noise current into noise voltage with = . Fig. 3e shows the corresponding maps. The average energy expenditure for the transduction of to was found to be in the range of = 160 fJ/pixel, which includes the read energy consumed by the in-memory WGN generator, calculated using Eq. 1.
Extended Data 6 shows the schematic and transfer characteristics of the MoS2 FET used as AN mimicking biological neurons with pre-synaptic voltage ( ) applied to the back-gate terminal   Therefore, the encryption can be considered to be secure from an eavesdropper with finite resources. Fig. 4f shows the average energy expenditure by a MoS2 AN for the encryption of the letter 'N" as a function of . Note that the energy expenditure is less than 10 pJ/pixel even for the highest .
The encryption strength is also tested assuming that the eavesdropper has access to a trained artificial neural network (ANN) and the information being communicated are encrypted MNIST data set for digit classification. Fig. 4g shows a fully connected two-layered ANN with 100 neurons in the hidden layer and 10 neurons in the output layer. The 10 output neurons correspond to digits from 0 to 9. MNIST images (28×28 pixels) are flattened to obtain corresponding 784×1 vectors, which are fed to the input layer. Gradient decent algorithm is used to train the ANN using 60,000 images with a learning rate of 0.001 and rectified linear unit (ReLU) as the activation function to ensure high convergence accuracy of 90.6% beyond 300 epochs. Following this, a testing accuracy of 92.2% was achieved using the remaining 10,000 images. Note that higher training and testing accuracies can be achieved by optimizing the network, which is not the primary focus of this work.
Next, we added white Gaussian noise to 10,000 MNIST images and binarized them at a threshold of 1.5 mimicking our MoS2 based artificial neural encoder. Fig. 4h shows some example of encoded MNIST images for different standard deviation ( ) of the WGN. Fig.4i shows the inference accuracy for the encrypted images as a function of , which follows a non-monotonic behavior. Interestingly, the accuracy values are found to be significantly low irrespective of , indicating the robustness of our proposed encryption scheme to trained ANNs.
In order to retrieve the information, we adopt population voting-based algorithm. We assume that  accurate decryption is found to be different for different . Therefore, not only and , but also prior knowledge of is required for decoding the information, which makes the system more robust from the eavesdropper.

Conclusion
In conclusion, we have experimentally demonstrated an all-in-one hardware IoT platform based on programmable and multifunctional MoS2 FETs, which is capable of sensing, storing, and securing information. Since a single material and similar device structures are used the hardware footprint is minimal, ~5 µm X 2 µm for each photodetector, ~5 µm X 3 µm for each CA, 5 µm X 64 µm for each WGN generator, and ~5 µm X 2 µm for each AN. The energy expenditure is also miniscule, in the range of few tens to hundreds of pico Joules, as we have primarily exploited subthreshold FET operation. Furthermore, analog, and non-volatile memory capability allows reconfiguration of the IoT platform based on application needs. Finally, the biomimetic IoT platform is shown to be secure.

Methods
Film Growth: Monolayer MoS2 was deposited on epi-ready 2" c-sapphire substrate by metalorganic chemical vapor deposition (MOCVD). An inductively heated graphite susceptor equipped with wafer rotation in a cold-wall horizontal reactor was used to achieve uniform monolayer deposition as previously described [22]. Molybdenum hexacarbonyl (Mo(CO)6) and hydrogen sulfide (H2S) were used as precursors.