Base station classifier methods and systems

Abstract

A base station classifier on a portable device includes code for obtaining base station resource allocations, converting the allocations into vectors to provide input to a neural network, and providing access to a trained neural network that evaluates the vectors to identify that a base station is a trusted vendor base station. A method of operations listens to resource allocations of a connected device, converts each allocation into a vector suitable for neural network analysis, evaluates the vector with a trained neural network with to determine if the vector is from a trusted base station, and connects only to base stations indicated by the trained neural network to be a trusted base station.

Description

INVENTORS: ALEXANDER MARDER; KIMBERLY CLAFFY; KA PUI MOK; AARON SHALEV; KYLE JAMIESON. BASE STATION CLASSIFIER METHODS AND SYSTEMS PRIORITY CLAIM AND REFERENCE TO RELATED APPLICATION

[0001] The application claims priority under 35 U.S.C. §119 and all applicable statutes and treaties from prior United States provisional application serial number 63 / 507,261, which was filed June 9, 2023. STATEMENT OF GOVERNMENT INTEREST

[0002] This invention was made with government support under 2226460 awarded by the National Science Foundation (NSF). The government has certain rights in the invention. TECHNICAL FIELD

[0003] Fields of the invention are cellular wireless communications and wireless device security. BACKGROUND

[0004] A base station interfaces with wireless phones and other cellular- enabled portable devices to provide access to various services, e.g., communications with traditional phone networks, other wireless networks and internet data access.Identification of base station vendor equipment when a portable device connects to a base station can provide benefits, e.g., enhanced services or security against possible back-door infiltration. Device security against infiltration from adverse base stations can provide important security to industry, governments and in conflict situations, including on or near battlefields.

[0005] Various government entities have considered excluding certain manufacturers from their infrastructure, and the present invention can provide a tool to ensure such policies are followed. As an example, Western officials have expressed concern with intelligence services in China placing backdoors in 4G and 5G base stations made by Chinese vendors. Such backdoors are intentional, long- lived, and provide adversaries with the ability to monitor, modify and exfiltrate traffic for offline analysis and future operations. Five base station vendors dominate the 5G market-Ericsson, Nokia, and Samsung, Huawei and ZTE.

[0006] Fingerprinting a radio transmission profile of different devices and manufacturers has been used to fingerprint phones. See, [7] Kennedy, et al. “Radio Transmitter Fingerprinting: A Steady State Frequency Domain Approach,” IEEE 68th Vehicular Technology Conference (2008), pp. 1-5; [6] Givehchian, et al, “Evaluating Physical-Layer BLE Location Tracking Attacks on Mobile Devices,” IEEE Security and Privacy (S&P) (2022), pp.1690–1704. However, transmission fingerprints are highly sensitive to changes in manufacturing and temperature, making it simultaneously more difficult to create comprehensive fingerprints, and easier for an adversary to circumvent our classifier. [5] Cekic, et al. “Wireless Finger-printing via Deep Learning: The Impact of Confounding Factors,” 55th Asilomar Conference on Signals, Systems, and Computers (2021), pp.677–684. SUMMARY OF THE INVENTION

[0007] A preferred embodiment provides a base station classifier on a portable device that includes code for obtaining base station resource allocations, converting the allocations into vectors to provide input to a neural network, and providing access to a trained neural network that evaluates the vectors to identifythat a base station is a trusted vendor base station. A method of operations listens to resource allocations of a connected device, converts each allocation into a vector suitable for neural network analysis, evaluates the vector with a trained neural network with to determine if the vector is from a trusted base station, and connects only to base stations indicated by the trained neural network to be a trusted base station. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] FIGs.1 and 2 show a preferred classifier that classifies base stations according to vendor by detecting how they allocate resources to connected devices. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] Preferred embodiments provide a passive classifier that can ID base station equipment without leaving a data fingerprint or establishing a data connection during the ID process, which can therefore prevent an unwanted connection if it is deemed to be undesirable. Such an ID process can be valuable for providing enhanced communications security, such as people employed at a competitive company or the government, either class of which can have enhanced security concerns. Preferred embodiment methods require only low-resource communications and analysis, and therefore can be easily run on a laptop with a portable USB-powered software defined radio, making it practical to deploy.

[0010] A security application of the present invention is the ability of an entity, such as a government entity, to detect potential unwanted entities, such as other government entities, and therefore placement of backdoors in 4G and 5G base stations.

[0011] A preferred method and device with code for implementing the method can operate through a 4G or 5G network and classify base station vendors to enable identification of threat status or to limit communication to use only base stations deemed to be trusted base stations. While all base stations appear the same to end user equipment, key implementation differences distinguish one vendor fromanother. A preferred classifier determines a base station's vendor in seconds by passively observing how it allocates resources to connected devices, without connecting to the base station or transmitting any data. The method, in addition to identifying known vendors, can also detect custom surveillance base stations whose allocation pattern does not match any known vendor. The classifier requires minimal computational resources and runs for hours without external power (e.g., on a laptop), facilitating testing and deployment. The classifier can be integrated with an external modem that connects a device to a specific 5G base station that the classifier identifies as trusted. Example data below shows data from a classifier that was trained and tested.

[0012] Two characteristics of commercial networks make the present classifier especially applicable to a government entity, because there is a high cost of acquiring, installing, and maintaining cell tower equipment that usually requires that networks operate a base station for years to get a return on investment, so frequent re-classification of the same base station is not necessary. Second, many networks use one base station vendor throughout a geographic region to simplify operations and ensure compatibility. In this common case, the problem of identifying trusted base stations reduces from selecting individual base stations to picking the right commercial network.

[0013] Preferred embodiments of the invention will now be discussed with respect to experiments and drawings. Broader aspects of the invention will be understood by artisans in view of the general knowledge in the art and the description of the experiments that follows.

[0014] FIGs. 1 and 2 show a preferred classifier 100 that classifies base stations 102 according to vendor by detecting how they allocate resources to connected devices 104. The classifier 100 decodes lightly encrypted broadcast allocations. The decoding technique is based upon the NG-Scope system of Yaxiong Xie available through the Internet on the git hub website. See also, Xie, Y., and Jamieson, K., “NG-Scope: Fine-grained telemetry for NextG cellularnetworks,” ACM SIGMETRICS 6,1 (Feb. 2022). NG-Scope is built on the SRS stack and runs on a laptop with a portable USB-powered software defined radio (SDR). Devices 104 can store a trained classifier 100 of the invention, or can communicate with a server that stores the trained classifier 100. An experimental trained classifier model used uses very little hard drive / SSD storage (<1 MB) to store the model and very little volatile memory (1-2 MB) during operation.

[0015] A preferred example classifier was trained and evaluated on base stations deployed on cell towers in AT&T, Verizon, and T-Mobile wireless networks to validate that the classifier 100 functions in real traffic and resource allocation profiles. We found ground truth using publicly available cell tower applications in Montgomery Country, MD and San Diego, CA. The classifier 100 includes a neural network and a prototype required only 11 seconds of observed allocations to distinguish Ericsson, Nokia, and Samsung base stations with perfect accuracy in our evaluation. Additional classifier training was conducted from data collection efforts that included base stations in other countries; e.g., Indonesia, China, Hong Kong, Spain, Italy, and Austria. This additional data collection has provided data to train a prototype classifier for all current large base station vendors - Ericsson, Nokia, Samsung, Huawei and ZTE. The neural network was trained with physical resource block allocations observed in the downlink control information channel. Training in different geographical areas with different carriers provided a good training data set as carriers have some degrees of freedom to tune parameters. Different countries typically have different carriers who can make different choices in tuning the parameters of a base station from a common manufacturer. Different traffic patterns can also be created by different user devices. For these reasons the neural network benefits from the training in different geographical areas.

[0016] The neural network was trained with allocations from 17 base stations and evaluated against 6 additional base stations on different cell towers than those used for training (Table 1). The classifier 100 correctly identified the vendors of allevaluation base stations, and it correctly classified every possible 11-second sequence of allocations across over 130 minutes of evaluation data. For 5 of the 6 evaluation base stations, the classifier 100 only needed up to six seconds of allocations to make perfect predictions. In FIG.2, a base-select module 202 acts as an external modem and selectively connects to a specific base station. The base- select module 202 locks a modem to a specific base station that is classified as acceptable by the classifier 100. A mobile device link layer 204 will select a base station based upon signal quality, but the base-select module 202 can veto a selection if the classifier determines that the base station is not trusted. A conventional modem + mobile device operating system picks a base station on the basis of signal quality, and the carrier network might also direct a modem to connect to a specific base station. The base select module 202 takes that decision out of the modem’s control (and the carrier’s control) if a base station is considered untrustworthy. Instead, the base select module locks onto a trustworthy cell even if it does not have the highest signal quality.

[0017] When base stations identified by the phone’s link layer 204 do not appear like any known trusted vendor, the classifier 100 identifies a low confidence match score indicating that it does not match the allocation profile of any known vendor, and the base-select module 202 directs the link layer 204 to use a different resource. The classifier 100 provides confidence scores of its correctness that can be used distinguish confident classifications from guesses. Rather than letting firmware (link layer 204 and operating system 206) independently select a base station, the classifier 100 identifies specific base stations as trusted for connection.

[0018] Base-Select technology used in the module 202 is implemented through the API provided by modem manufacturers. Some manufacturers (e.g., Quectel) include a specific instruction to lock to a cell or set of cells that can connect to a specific base station using the cell identifier number. The modified base-Select module 202 that works with the classifier 100, uses output from the classifier to select the SIM card necessary to connect to the specified base station.

[0019] While a classifier 100 was implemented with a software defined radio (SDR) and a laptop, the same exact process could occur entirely within a phone or other wireless device. The following are preferred steps for the classifier 100 to implement for training purposes, which steps are also used by a portable device 104 during operation to identify trusted base stations to allow a connection.

[0020] Step 1: Tune Radio to Base Station Frequency. Learning a base station's downlink frequency can occur in many ways. As an example, equipment can be placed near a cell tower and a phone used to learn the EARFCN (E-UTRA Absolute Radio Frequency Channel Number) of the strongest signal. A radio is then tuned to that frequency.

[0021] Step 2: Listen to Base Station Resource Allocations. Once the radio is tuned to the base station's frequency, resource allocations to connected devices can be discovered via decryption. In each subframe, the base station broadcasts which devices, identified by a Radio Network Temporary Identifier (RNTI), will receive a physical resource block (PRB) allocation, along with the modulation coding scheme (MCS), the transmission block size (TBS), and other information. These allocations are extracted, while ignoring the data portion of the radio transmissions, to derive a schedule of allocations.

[0022] The following table shows an example allocation to a device with RNTI (radio network temporary identifier) 100. RNTI PRBs MCS 1 MCS 2 TBS 1 TBS 2 100 72 13 4 28336 9912

[0023] That RNTI is converted by the classifier 100 into a normalized vector that is useful for neural network analysis, as shown in the following table. 0.72 0.42 0.13 0.29 0.10

[0024] Step 3: Prepare Allocation Schedule for Deep Learning. Each vector contains the PRBs, MCS indexes, and TBSes. Neural networks prefer normalized inputs, so MCS and TBS values are normalized for training and analysis according to 3GPP (3rdGeneration Partnership Project) specifications, and the PRBs arenormalized based on the maximum number of PRBs in the frequency band. Normalizing the PRBs in this fashion allows multiple allocations to be represented per subframe and reduces the impact of the inherent bandwidth differences between frequency bands.

[0025] Since a base station might make multiple allocations in a single subframe, each subframe vector contains 100 cells, sufficient for 20 device allocations. The resulting vector for a subframe is the concatenation of each allocation vector, in reverse order of PRB allocation size. The following table shows an example of the original vectors. Each of vectors corresponding to subframes 4220-4224 and 4225-4229 will be combined into a single vector so that each vector includes 5 subframes. Combining the vectors is accomplished by appending each vector to the end of the prior vector.

[0026]

[0027] bins containing 5 consecutive subframes each; i.e., bin0-4contains subframes ^0െ 4^ while bin5-9contains the next five subframes. Each bin is then coverted to a single vector as shown in the following example:

[0028] which is thethat the resulting vector has 500 cells. Five subframes per vector is preferred toincrease the patterns available to the neural network compared to smaller vectors, without allowing the neural network to overfit the patterns to the training data.

[0029] Step 4: Training the Neural Network. A prototype neural network had 2 hidden layers, and 50 neurons per layer. Training with different number of layers and neurons produced similar results, although fewer layers tended to underfit while more layers tended to overfit to the training data. Higher numbers of neurons were tested, and training took longer without any accuracy benefit. However, larger training sets might benefit from increasing the number of layers and / or neurons to recognize sufficiently rare patterns.

[0030] Additional information broadcast along with the PRB allocations can be added to strengthen training. The following allocations are available to train a classifier of the invention.

[0031] Feature Description

[0032] RV 1 - Redundancy version 1

[0033] RV 2 - Redundancy version 2

[0034] HARQ - HARQ (Hybrid ARQ(Automatic Repeat reQuest) index

[0035] NDI 1 - Indicates retransmission

[0036] NDI 2 - Indicates retransmission

[0037] C-RNTI - Indicates a change in RNTI identifier for a UE (customer user equipment that connects to the cellular network; e.g., phones and tablets)

[0038] Paging - Identifies that the data is related to cell paging

[0039] A sparse matrix stores the normalized vector and can be used during training to reduce memory demands. Instead of using a dense matrix, where zeros are explicitly represented, a sparse matrix that only explicitly represents non-zero values is used. This does not impact accuracy but does decrease the memory and time required to train. Zeros occur when there is no allocation or false value.

[0040] For example, an original dense matrix with explicit zero values is shown below.0 0 0 0 0 .5 0 0 0 0 0 .3 0 0 0

[0041] The above matrix can be reduced, and the resulting sparse matrix reducing the required storage from 15 to 3 × 2 = 6 values. In experiments, training data was collected that would exhaust 500 GB of RAM when put into a dense matrix, which included about 395.2 million physical resource block allocations. A sparse matrix reduced the training data storage to around 92 GB.

[0042] Step 5: Predict Base Station Vendor. The trained classifier can predict the vendor for a base station, provided the vendor was part of training, while a low confidence prediction can indicate an unknown / untrusted base station. Data is processed by the classifier using steps 1 – 3 above. The model in the classifier also outputs a confidence value, allowing a mobile device to determine if the prediction should be trusted, which can be based upon the level of desired threat protection configured in the mobile device, which can be accessed through a menu, for example, or set by an administrator when mobile devices are issued or used by persons in a company, agency, or military unit, for example. The confidence value is especially useful when the vendor is not included in training; e.g., surveillance / rogue base stations.

[0043] A recurrent neural network (RNN) is used in a preferred classifier. RNNs are better for time-series data, which is the type of data contained in the schedule of allocations. Unlike other neural networks, RNN predictions are based on the current vector (during both training and operation) as well as feature vectors seen previously. Additional experimentation trained an RNN where each normalized vector represents a single subframe, but the RNN can take advantage of prior information for accurate predictions. The model tested used one recurrent layer and one conventional layer. Specifically, data was represented as one vector per subframe (rather than the 5 subframe vectors used with the conventional neural network), and then grouped subframes into contiguous buckets. The RRN ispresented with an entire in-order bucket. All predictions (training and operation) make a prediction for the last vector in the bucket, and this prediction is influenced by the vectors that preceded it in the bucket. As an example, groups of 100 subframe vectors can be presented to the RNN.

[0044] While specific embodiments of the present invention have been shown and described, it should be understood that other modifications, substitutions and alternatives are apparent to one of ordinary skill in the art. Such modifications, substitutions and alternatives can be made without departing from the spirit and scope of the invention, which should be determined from the appended claims.

[0045] Various features of the invention are set forth in the appended claims.

[0046]

[0047] References

[0048] [1] NG-Scope. https: / / github.com / YaxiongXiePrinceton / NG-Scope.

[0049] [2] QXDM Professional Tool Quick Start. https: / / www.qualcomm com / content / dam / qcomm-martech / dm-assets / documents / 80- n94711_d_qxdm_professional_tool_quick_start.pdf.

[0050] [3] Balasingam, A., Bansal, M., Misra, R., Nagaraj, K., Tandra, R., Katti, S., AND Schulman, A. Detecting if LTE is the Bottleneck with BurstTracker. In ACM MobiCom (2019), pp.1-15.

[0051] [4] Bui, N., AND WidmER, J. OWL: A Reliable Online Watcher for LTE Control Channel Measurements. In ACM AllThingsCellular (2016), pp. 25- 30.

[0052] [5] Cekic, M., Gopalakrishnan, S., And Madhow, U. Wireless Fingerprinting via Deep Learning: The Impact of Confounding Factors. In 2021 55th Asilomar Conference on Signals, Systems, and Computers (2021), pp. 677- 684.

[0053] [6] Givehchian, H., Bhaskar, N., Herrera, E. R., Soto, H. R. L., Dameff, C., Bharadia, D., AND Schulman, A. Evaluating Physical-Layer BLELocation Tracking Attacks on Mobile Devices. In IEEE Security and Privacy (S&P) (2022), pp.1690-1704.

[0054] [7] Kennedy, I. O., ScAnlon, P., Mullany, F. J., Buddhikot, M. M., Nolan, K. E., AND RONDEAU, T. W. Radio Transmitter Fingerprinting: A Steady State Frequency Domain Approach. In 2008 IEEE 68th Vehicular Technology Conference (2008), IEEE, pp.1-5.

[0055] [8] Kumar, S., Hamed, E., Katabi, D., And Erran Li, L. LTE radio analytics made easy and accessible. ACM SIGCOMM Computer Communication Review (CCR) 44, 4 (2014), 211-222.

[0056] [9] Li, Y., Peng, C., YuAn, Z., Li, J., Deng, H., And WAng, T. MobileInsight: Extracting and Analyzing Cellular Network Information on Smartphones. In ACM MobiCom (2016), pp.202-215.

[0057]

[0010] PAncevski, B. U.S. Officials Say Huawei Can Covertly Access Telecom Networks. https: / / www.wsj.com / articles / u-s-officials-say-huaweican- covertly-access-telecom-networks-11581452256, 2020.

[0058] Systems. Open Source srsRAN project. https:

[0059]

[0012] Trendforce. Three companies to own 74.5% of base station market in 2022. https: / / evertiq.com / news / 52267.

[0060]

[0013] WAN, T. False Base Station or IMSI Catcher: What You Need to Know. https: / / www.cablelabs.com / blog / false-base-station-or-imsicatcher-what- you-need-to-know, 2019.

[0061]

[0014] XIE, Y., AND JAMIEson, K. NG-Scope: Fine-grained telemetry for NextG cellular networks. ACM SIGMETRICS 6, 1 (Feb.2022).

Claims

CLAIMS 1. A base station classifier on a portable device, the classifier comprising code for obtaining base station resource allocations, converting the allocations into vectors to provide input to a neural network, and providing access to a trained neural network that evaluates the vectors to identify that a base station is a trusted vendor base station.

2. The base station classifier of claim 1, further comprising code to limit connection of the portable device to base stations that have a predetermined confidence level of being trusted.

3. The base station classifier of any previous claim, wherein the trained neural network is stored on the portable device.

4. The base station classifier of any previous claim, wherein the trained neural network is accessed via a server.

5. The base station classifier of any previous claim, wherein the trained neural network comprises a recurrent neural network.

6. The base station classifier of claim 1, wherein the code comprises code to prevent a link layer and operating system of the portable device from independently selecting a base station.

7. The base classifier of claim 6, wherein the code for converting includes code to decrypt base station resource allocation broadcast information.

8. The base classifier of claim 7, wherein the code for converting decrypts Radio Network Temporary Identifier (RNTI), physical resource blocks (PRBs), modulation coding scheme (MCS), and transmission block size (TBS) from a base station resource allocation subframe.

9. The base classifier of claim 7, wherein the code for converting decrypts one or more of RV 1 - Redundancy version 1, RV 2 - Redundancy version 2, HARQ - HARQ (Hybrid ARQ(Automatic Repeat reQuest) index, NDI 1 - Indicates retransmission, NDI 2 - Indicates retransmission, C-RNTI - Indicates a change in RNTI identifier for a UE, Paging - Identifies that the data is related to cell paging.

10. The base classifier of claim 6, wherein the code for converting converts the allocations into a sparse matrix of vectors.

11. The base station classifier of claim 1, wherein the trained neural network outputs a confidence score.

12. A method for base station classification on a portable device, the method comprising: with the portable device, listening to resource allocations of a connected device; converting each allocation into a vector suitable for neural network analysis; evaluating the vector with a trained neural network with to determine if the vector is from a trusted base station; and connecting only to base stations indicated by the trained neural network to be a trusted base station.

13. The method of claim 12, wherein the evaluating includes the trained neural network providing a confidence level.

Images