In our paper, we introduce a universal web service user's identification method. This method is based on analyzing the digital fingerprint of the visitor using a neural network. Within the scope of our research, we performed a comparative analysis between our developed method and the existing fingerprint detection services. The testing results indicate that the accuracy of fingerprint identification using our method surpasses fingerprint.com by 3.1% on desktop platforms and 6.3% on mobile devices. Furthermore, the utilization of our method significantly reduces the number of false positive errors, thereby enhancing the robustness of user identification against variations in browser and device parameters.
identification.
Browser fingerprint or device fingerprint, combined into the concept of a digital fingerprint, is information collected about the software and hardware of a remote device for the purpose of its
The technique of digital fingerprinting has existed for many years. The first mentions of various techniques for obtaining and analyzing digital fingerprints in scientific literature appeared in 2003 [4], and they have been widely studied since 2009 [5].
Since then, many different techniques for determining the digital fingerprint have been described: • • • • • • • •
JavaScript-Based Fingerprints CSS-Based Fingerprints Canvas-Based Fingerprint Hardware and Software-Based Fingerprints Fingerprint Based on Audio API Plugin-Based Fingerprint TLS Fingerprint Other Browser Fingerprint Acquisition Technologies (correlation between visitor's gaze and mouse movement; characteristics of HTML parser; font sets (font glyphs); methods based on calculation of JavaScript scripts set execution time; based on user lag time on websites; on the nature of user interaction with touchpad; speed and specificity of typing on keyboard; speed and directions of mouse movement).
In most cases, to identify a digital fingerprint, a scheme is used in which code based on a special JS library is executed on the client side. The code performs a set of tests and checks defined by the library and send the received parameters to the server. Usually, the server is deployed as a separate service (Figure 1).
All modern methods of identifying digital fingerprinting have both advantages and disadvantages.
The main drawbacks of fingerprinting solutions include: • • •
Low user identification accuracy, Computation time for generating a digital fingerprint,
Time required for matching with previously known digital fingerprints in the system, • • •
Short lifespan of a specific digital fingerprint, High device load on the user's end, Dependence on JavaScript, Challenges in computing a digital fingerprint in homogeneous environments (computer labs, internet cafes, mobile network environments), Cross-browser digital fingerprinting, Low accuracy in identifying users operating in incognito mode,
Matching digital fingerprints over VPN.
In addition to the mentioned drawbacks of existing methods for digital fingerprinting based on open solutions, ready-made commercial services are characterized by additional disadvantages:
High cost, Closed source nature, Data stored on third-party servers,
Dependence on the service provider.
In our assessment, there are currently no effective methods that reliably identify a user based on their digital fingerprint over an extended period, especially when using VPN, incognito mode, or engaging in cross-browser surfing.
We reviewed some research papers that address the problems of fingerprinting and user identification on the Internet.
In [6], the authors reviewed and classified the existing fingerprinting techniques and their applications for user identification on the Internet and analyzed in detail the development of different research directions of browser fingerprinting. Based on the analysis of existing results, the problems faced by different research directions are pointed out. Also, the research achievements in the field of browser fingerprint recognition are summarized and the trend of future development is pointed out. The authors also discussed the privacy issues associated with the use of fingerprinting techniques.
The authors of the paper [7] show that GPU information obtained using WebGL and other technologies can be used to create a unique device fingerprint that can be used for user identification. At the same time, the authors note that changing GPU settings and parameters can change the device fingerprint, which makes identification more difficult.
In the study [8], the authors demonstrate the correlation between gaze and mouse movements and argue that this serves as a valuable source for obtaining browser fingerprints. Simultaneously, the authors point out that collecting data on a person's gaze in the browser has drawbacks, such as inaccuracies when using a webcam and the limitation that users must grant permission for camera access. The study also reveals that, in the case of computers used by multiple users, browser statistics may malfunction and can no longer differentiate between individuals.
In the article [9] authors analyze the popularity of the Transport Layer Security (TLS) protocol on the Internet and its use in censorship circumvention tools. The researchers collected and analyzed a huge volume of real-world TLS traffic to identify the different implementations of TLS clients used on the Internet. Censors can use deep packet inspection (DPI) to identify and block such tools based on their TLS fingerprints. That said, many circumvention tools fail to properly mimic popular TLS implementations, leading to their detection and blocking. To solve the censorship circumvention problem, the authors proposed a solution that allows developers to automatically mimic other popular TLS implementations. Using real-world data, the authors of the paper propose methods to flexibly adapt TLS-fingerprint to the dynamic TLS ecosystem with minimal manual effort.
The authors of the paper [10] propose a new mobile device user's identification method based on the study and analysis of touch dynamics, which has stable patterns of interaction between the user and his mobile device, including factors such as touch force, swipe speed and duration of touch.
This method has shown excellent results, but its scope is limited to only a subset of mobile devices and depends on the availability of APIs for interacting with physical device elements.
In the paper [11], the authors propose a browser fingerprinting defense tool to anonymize users' browsers. The authors show that browser fingerprinting cannot be prevented by the user. Although new methods are constantly being developed that can prevent browser fingerprinting, they cannot prevent it completely.
In the article [12], presents new algorithms for encoding and comparing fingerprints, which focus on the values of parameters with low stability and low entropy.
The method proposed by the authors allows for:
Improved accuracy in user identification under specified conditions, Reduce the percentage of false positives, Increased lifespan of the calculated digital fingerprint,
Maintenance of the speed of digital fingerprint identification at an industry-standard level.
All of these improvements are achieved through the implementation of a novel neural network training algorithm. The results of determining the digital fingerprint of a web service user are a nonlinear time series consisting of a set of browser and user device parameters and may vary over time [13, 14]. As the practice of the last 10 years shows, recurrent neural networks (RNN) are the most effective architecture for solving time series problems that cannot be solved by feedforward networks [15]. We performed comparative tests of the two most common RNN architectures LSTM and GRU by the methodology described in [16]. The results of the digital fingerprint accuracy tests are presented in Figure 2.
For our solution, we utilized the LSTM architecture as it demonstrated significantly better results over a small number of training epochs (50-100 epochs). This implies that, with equal resource consumption, LSTM yields superior results, which can be expressed by: → ! → " #.
(
Currently, the majorities of systems for obtaining a digital fingerprint are based on the fingerprint.js library or incorporate some of its functions. This library, one of the earliest to emerge, is dynamically evolving and includes prospective developments that emerge periodically. The library is actively developing, and the project repository is frequently updated. As of December 2023, the latest version is 4 [17]. Starting from this version, the developer has changed the distribution terms, and it is now offered under the Business Source License 1.1. Currently, the FingerprintJS service is considered an industry standard.
The service allows for the identification of numerous browser and operating system parameters. The key modules of the fingerprint.js library are outlined in Table 2. Color inverted mode areColorsInverted() Colors forced
areColorsForced() Monochrome depth getMonochromeDepth() Reduced motion
isMotionReduced() Video card (WebGL) getVideoCard() Contrast HDR Math calc Font width PDF viewer Architecture getContrastPreference() isHDR() getMathFingerprint() getFontPreferences() isPdfViewerEnabled() getArchitecture() boolean | undefined number | undefined number | undefined
The general algorithm of operation for the fingerprint.js library is presented in Figure 3.
At the initial stage of preparing data for training the neural network, we have a multidimensional dataset about the user collected in the previous stage. To optimize time and computational resource costs, this multidimensional dataset is transformed into a linear vector. Thus, the neural network receives a one-dimensional vector as input.
Next, after normalization, the data is randomly split into testing and training sets in a 30%/70% ratio.
Based on the testing set, a prediction is made to determine if the visitor is known in our service, and the prediction result is compared with the result obtained based on the predefined parameters of the model. The schematic process of training the neural network is illustrated in Figure 4.
The initial training of the model was conducted using the "Login Data Set for Risk-Based Authentication" dataset from Kaggle [13]. This dataset includes a list of parameters associated with each login attempt.
The structure of the dataset is presented in Table 2. HTTP Accept Headers Language Screen Resolution Timezone Browser Plugins Platform (Operating System) Browser Version Device Memory Canvas Fingerprint WebGL Vendor and Renderer Cookies Enabled
Type String String String Integer (Width x Height) String List of Strings [] String String Integer (in gigabytes) String (hashed or raw data) String
Range or example Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/58.0.3029.110 Safari/537.36.
Access-Control-Allow-Origin: *, Cache-Control: max-age=604800, Content-Type: multipart/form-data, If-Unmodified-Since: Mon, 27 Nov 2023 12:43:00 EET uk 2073600 Europe/Kiev Linux x86_64 Chrome 119 8 [PDF Viewer, Chrome PDF Viewer, Chromium PDF Viewer, Microsoft Edge PDF Viewer, WebKit built-in PDF] 93a13b9b08d18393f5c731f8f5c58a11 WebKit WebGL
TRUE Boolean FALSE ["4274,142 default, cursive, fantasy", "4314,143 sans-serif, Arial, Arimo, Helvetica, Liberation Sans", "4249,142 serif", "3780,149 monospace", "4431,143 system-ui, Ubuntu", "4189,143 aakar"] 13b9b08d18393f5c731f8f5c58a116dcb 4 FALSE FALSE 4g FALSE TRUE 4 Desktop
To perform an experiment comparing the effectiveness of the developed method and the method of digital fingerprinting using the FingerprintJS service, a set of parameters from 2134 devices of different types (desktop computers, mobile devices, tablets) and a set of user agents that was generated using the npm package User-Agents [18] were used. User-Agents are a JavaScript package for generating random user agents based on how often they are used in a real environment.
The generated data includes hard-to-find browser fingerprint properties, and powerful filtering capabilities allow the generated user agents to be constrained to fit specific needs.
An experiment to measure the qualitative performance of the developed web service user identification method was performed on the current web service using the algorithm that is shown in Figure 5.
The results of the experiment are summarized in Tables 3 – 5.
The accuracy comparison data for digital fingerprint identification indicate that for desktop computers, the accuracy of the existing identification method (FingerprintJS) is 90.7%, while the accuracy of our developed method is 93.7%, representing a 3.1% improvement.
For mobile devices, the accuracy of the existing user identification method (FingerprintJS) is 89.7%, whereas the accuracy of our developed method is 96%, showcasing an improvement of 6.3%.
In the case of tablets, the accuracy of the existing identification method (FingerprintJS) is 89.2%, which is 5.4% lower than that of our developed method (94.6%).
The weighted average accuracy of the method developed by us is 3.8% higher than the existing method (94.2% versus 90.4%).
The stability of the algorithm directly depends on reducing the percentage of false positives and
false negatives in user identification. The stability of the algorithm can be determined using equation
$ %&'& () + (+ , (
() + (+ + ,) + ,+ where TP - true positive, TN - true negative, FP - false positive, FN - false negative.
The method developed by us shows a lower number of false positive fingerprint identification results on all investigated platforms: • • •
Desktop computers: 52 versus 105, Mobile devices: 8 versus 26,
Tablets: 3 versus 7. • • •
Desktop computers: 49 versus 47, Mobile devices: 7 versus 16,
Tablets: 3 versus 4.
The weighted average number of false positive errors for the developed method is 41.0, compared to 84.9 for the existing method.
The number of false negative results in digital fingerprint identification is comparable for both methods on all investigated platforms, with the advantage of the developed method being notably better only on mobile devices:
The weighted average number of false negative errors for the developed method is 38.6, compared
to 38.9 for the existing method. According to formula (
The duration of the identification process using the developed method varies in the ranges of 5977 ms for desktop computers, 143-181 ms for mobile devices, and 92-110 ms for tablets. Based on the comparison results, it can be concluded that the speed of user identification using the developed method is comparable to the speed of identification using existing modern methods.
The analysis of the obtained results shows that the developed method has higher accuracy on all investigated types of devices and platforms. Additionally, it exhibited a lower overall error rate in the accuracy of identification and comparable speed in the process of digital fingerprint determination.
The author(s) have not employed any Generative AI tools. [6] D. Zhang, J. Zhang, Y. Bu, B. Chen, C. Sun, T. Wang, A survey of browser fingerprint research and application, Wireless Communications and Mobile Computing, 2022. doi: 10.1155/2022/3363335. [7] DRAWNAPART: A Device Identification Technique based on Remote GPU Fingerprinting, 2022.
URL: https://arxiv.org/abs/2101.03793. [8] W. Fuhl, N. I. Sanamrad, E. Kasneci, The Gaze and Mouse. Signal as additional Source for User
Fingerprints in Browser, 2022. URL: https://arxiv.org/abs/2101.03793. [9] E. Wustrow, S. Frolov, (University of Colorado Boulder), The use of TLS in Censorship
Circumvention, doi:10.14722/ndss.2019.23511.
[10] B. Pelto, M. Vanamala, R. Dave, Your Identity is Your Behavior -- Continuous User
Authentication based on Machine Learning and Touch Dynamics, 2022. URL:
https://arxiv.org/abs/2305.09482.
[11] D. Moad, V. Sihag, G. Choudhary, Fingerprint defender: Defense against browser-based user
tracking. In: I. You, H. Kim, TY., Youn, F. Palmieri, I. Kotenko (Eds.), Mobile Internet Security.
MobiSec 2021, volume 1544 of Communications in Computer and Information Science, Springer,
Singapore, 2021. doi: 10.1007/978-981-16-9576-6_17.
[12] M. Gabryel, K. Grzanek, Y. Hayashi, Browser Fingerprint Coding Methods Increasing the
Effectiveness of User Identification in the Web Traffic, Journal of Artificial Intelligence and Soft
Computing Research 10(
Choosingthe best Architecture for Passenger Traffic Data. Automation and computer-integrated technologies 2(72) (2022) 38–44. doi: 10.18372/1990-5548.72.16941. [17] GitHub, fingerprintjs/fingerprintjs: Browser fingerprinting library, 2023. URL: https://github.com/fingerprintjs/fingerprintjs. [18] User Agents, 2022. URL: https://www.npmjs.com/package/user-agents.