Fingerprinting light emitting diodes using spectrometer

Fingerprinting refers to a class of techniques to distinguish products using their unique features originated during production. Fingerprinting has a wide range of applications including supply- chain integrity and device authentication. The authors propose the ﬁ rst ﬁ ngerprinting technique of light-emitting diode (LED), which is pervasively used in electronics products. The key idea is to use spectral features of LEDs obtained by a spectrometer. Combined with a machine-learning classi ﬁ er, the proposed method successfully dis-tinguishes ten individual LED samples from the same lot at 99% accuracy.

Proposed method: In this Letter, we consider white LEDs that are commonly used for lighting. Among many ways to realise white light, combining a blue LED and yellow phosphor material is the most common for its cost efficiency [5] (see Fig. 1). The phosphor material absorbs blue light from LED and emits yellow light by photoluminescence. Human eyes recognise the mixture of the complementary colours, blue and yellow, as (pseudo) white. This type of LED is called phosphor-converted white LED.
The colour of a white LED is susceptible to a slight manufacturing process variation. To tackle the problem, the technique called LED binning comes into play: the colour of an LED is projected to the CIE 1931 chromaticity diagram which is separated into small bins. A manufacturer ensures that a particular set of LEDs fits within a specific bin.
The above discussion suggests that each LED can have a unique colour, but that is within a small bin. A slight variation should be detected to find the difference between LEDs. To address the problem, we propose to use optical spectrum for fingerprinting LED. The optical spectrum is a decomposition of light into each wavelength and contains more information which is degenerated when it is projected to a chromaticity diagram. Fig. 1 shows a typical spectrum from the phosphor-converted LED: the blue light from the LED makes a sharp peak around 450 nm, while the yellow light from the phosphor material makes a relatively large bump around 600 nm. The proposed fingerprinting uses a coarse-to-fine strategy. In the first step, we distinguish samples by their part numbers. In the second step, we distinguish a particular sample from many samples of the same part number.

Experiments
Setup and preprocessing: Fig. 2 shows the experimental setup. A stabilised power supply drives the LED, and the light from the LED is captured by a spectrometer from 47.5 cm away. The LED holder and the spectrometer are fixed to a base plate for the reproducibility of experiments. A shielding box covers the entire setup during the measurement. The DC power supply drives the target LED sample with the constant current of 20 mA.  The Hamamatsu C12666MA MEMS spectrometer [7] is used. For each measurement, we get a 256-dimensional vector [s(0), . . . , s(255)] that represents the spectrum between 340 and 780 nm wavelength at 256 equal steps. The low-cost spectrometer is composed of a grating and a linear image sensor and is available with a few hundred dollars. Table 1 summarises the target LEDs. We use ten samples for three different LED products from different manufacturers, and they are referred to as L A , L B , and L C . As a preprocessing, the total power of measured spectra, which is susceptible to measurement environment, is normalised: the 256dimensional vector [s ′ (0), . . . , s ′ (255)] that represents the normalised spectrum is obtained as Fingerprinting LEDs with different parts numbers: As a first step, we distinguish LEDs by a part number. Fig. 3 shows the normalised spectra in (1) from the 30 samples summarised in Table 1. The horizontal and vertical axes are the wavelength and the normalised light intensity, respectively. There are 30 traces, and their colour corresponds to their part number: blue, red, and green traces represent L A , L B , and L C , respectively.
wavelength, nm   3 shows that there is a distinct difference between L a , L B , and L C in the yellow region (500-600 nm). There is a particularly large difference around 500 nm which is a boundary between the blue and yellow spectra. Therefore, we can distinguish LEDs by a part number by simple thresholding. The difference can be attributed to different semiconductor processes and phosphor materials.
Fingerprinting individual LED samples: As a second step, we distinguish a specific LED sample from others with the same part number. In this particular experiment, we examine L A . To check the reproducibility of measurements, each sample is measured for ten times.
To emphasise the small difference, we subtract the total mean from the measured spectra. We describe the normalised spectrum of the kth measurement of the ith sample as (2) Fig. 4 shows the normalised spectra defined in (2) for 100 traces: 10 measurements of each of the 10 samples. The region higher than 550 nm is omitted because no significant difference is observed. The traces are coloured by samples. The figure shows that each sample makes clusters. These traces suggest that we can distinguish each sample by using an appropriate distinguisher. Each sample is measured ten times. The traces are coloured by samples Distinguishing samples using machine-learning classifiers: To evaluate the uniqueness of each sample quantitatively, we classify the ten samples in the previous experiment using classification algorithms. For the purpose, the measured data is classified using nine representative supervised-learning algorithms in the scikit-learn machine-learning library. Table 2 shows the performance figures of the 10-class classifiers. In Table 2, the success rate refers to the one obtained by averaging the success rates of the leave-20%-out cross-validation. LogisticRegression and KneighborsClassifier show the best results with 99 and 98% success rates. The results show that the individual LED samples are distinguishable at a high success rate with an appropriate classifier.

Conclusion:
We proposed the first fingerprinting technique of white LED using spectral features. The feasibility of the method is verified through concrete experiments. LEDs having different part numbers can be easily distinguished attributed to the difference in LED fabrication processes and phosphor materials. Individual samples of the same part number are still different in the blue region. By using a machinelearning classifier, ten individual samples can be distinguished at the 99% success rate. This Letter only covers the feasibility study, and there are remaining works before practical application. In particular, we plan to study (i) additional experiments with a larger number of samples and (ii) performance evaluation as a realistic authentication system.