Flow-Based Detection and Identification of Zero-Day IoT Cameras
Priyanka Rushikesh Chaudhary, Rajib Ranjan Maiti · 2025
Read Flow-Based Detection and Identification of Zero-Day IoT Cameras by Priyanka Rushikesh Chaudhary, Rajib Ranjan Maiti (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-08485.txt.
Introduction
Unknown cameras joining a network announce themselves in their traffic, and Priyanka Rushikesh Chaudhary with Rajib Ranjan Maiti build a detector in a 2025 network security paper. Consumer internet-of-things devices rarely give administrators monitoring or control, so simply noticing a new streaming camera is a security problem in itself.
Their zCamInspector identifies known cameras with supervised classifiers and flags zero-day cameras, models never seen before, with one-class classifiers trained on normal patterns. Roughly forty gigabytes of traffic across three datasets, commercial cameras, open-source cameras, and conferencing applications as decoys, support the evaluation. Homes and offices full of silent lenses need exactly this kind of flow-based census before any policy can be enforced.
Catalog record
| CREATOR | Priyanka Rushikesh Chaudhary, Rajib Ranjan Maiti |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 75 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-08485
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Flow-Based Detection and Identification of Zero-Day IoT Cameras Priyanka Rushikesh Chaudhary, Rajib Ranjan Maiti Abstract The majority of consumer Internet of Things (IoT) devices do not provide explicit mechanisms for network administrators to manage, monitor, and control them, which hinders the deployment of tailored IoT security policies. In particular, an administrator may notice, possibly by monitoring a range of IP addresses, that a new device has initiated its communication, but it is difficult to decide if it is an IoT device, in particular, a streaming IoT camera. In this paper, we design and develop a new system, called zCamInspector , to identify a known set of IoT cameras using efficient supervised classifiers, called zCamClassifier, and to detect zero-day IoT cameras using efficient one-class classifiers, called zCamDetector. We have considered a total of about 40GB of network traffic categorized into three different broad types, Set I - consisting of six commercially available IoT cameras from different manufacturers - collected from our own IoT laboratory, Set II - consisting of five IoT cameras with about 1.5GB of open-sourced dataset, and Set III - consisting of four online audio/video conferencing applications and two video sharing applications - indicating non-IoT camera traffic. A set of 62 generic flow-based features are extracted with the CICFlowmeter tool. Using the ExtraTree classifier, we show that top 10 features in these three datasets have a number of common features. zCamInspector uses seven supervised ML models, namely ET, DT, RF, KNN, XGB, LKSVM, and GNB, and four one class classifiers, namely OCSVM, SGDOCSVM, IF,and DeepSVDD. Our results show that zCamInspector achieves the highest accuracy of more than 99% to identify the IoT cameras using XGB and false negatives as low as 0.3% in both datasets. Compared to the state-of-the-art, zCamInspector using XGB achieves a better precision to identify some of the IoT cameras in both Set I and Set II. In detecting zero-day IoT cameras, our results show that accuracies are 93.20%, 96.55%, 78.65% and 92.16% using OneClassSVM, SGDOneClassSVM, IsolationForest,and DeepSVDD respectively, using trained samples of Set III (i.e., Others data set) and tested on test samples of Set I and Set II (i.e., BITSHPC and UNSW data sets). We demonstrated the system’s effectiveness in detecting zero-day IoT cameras under scenarios where all 11 devices are previously unseen, and when only one IoT camera is zero-day and rest are part of the network. When all the IoT cameras become zero-day, DeepSVDD (mean training accuracy as 96.03% and mean testing accuracy as 74.51%), showing dominating performance over remaining three models OneClassSVM (mean training accuracy as 85.64% and mean testing accuracy as 60% ), IsolationForest (mean training accuracy as about 89.75% and mean testing accuracy as 62.79%) and SGDOneClassSVM (mean training accuracy of about 43.21% and mean testing accuracy as 67.05%). Comparing with state-of-the-art, the detection of zero-day IoT cameras is shown for the first time in this paper and the results show good performance of zCamInspector by achieving more than 95% accuracy in case of certain cameras, like Spy Clock camera. I Introduction Several recent market studies have revealed that the usage and prices of indoor and/or outdoor IoT surveillance cameras will occupy a huge volume that is worth 13.2 billion USD by 2032 [ 1 , 2 ] . According to NIST [ 3 , 4 ] , a vast majority of IoT devices cannot be accessed, managed, and monitored in the same way as conventional IoT devices, like Desktop, Laptop, Printers, Smartphones, and Tablets. This is because the IoT devices are often considered as “black boxes” to the network administrators, having no information about their states, meaningful identities, or the external systems they interact with. This is underpinned by the fact that rule-based attack detection mechanisms [ 5 , 6 , 7 , 8 ] are challenged due to the adoption of a same set of cloud service providers by a larger number of IoT vendors [ 9 ] . Among others, IoT cameras pose direct threats to user privacy where face images, behaviors, gestures, personally identifiable information (PII), phone numbers, home addresses, or home assets can be directly exposed to attackers [ 10 ] . Certain case studies on specific IoT cameras (like VibeX Onvif YY HD IoT camera [ 11 ] and Belkin NetCam camera [ 12 ] ) and IoT Hub (like SmartThingsHub [ 13 ] that can integrate IoT cameras) reveal that the risk of security breach is higher in IoT cameras compared to other IoT devices. Due to a lack of standardized mechanisms, the network administrators face humongous challenges to manage these devices using a centralized management system and this is aggravated by the scale of IoT devices in the network. In this paper, we address the problem of managing such IoT cameras by analyzing and identifying their network flows without relying on any IP address or transport ports or breaking encryption. Fig. 1 : A motivating example. We motivate our study by using a smart home scenario (Fig. 1 ), similar to that as envisioned in [ 14 , 15 ] . The smart home has a number of traditional computing devices, like a desktop or a smartphone, running certain online audio/video conferencing applications like Meet and video sharing applications like YouTube. Note that these applications do generate video flows, and these flows are acceptable usually. Also, certain IoT devices like smart bulbs and smart doorbells may be used in the environment. However, installation of a multi-functional IoT device, like smart bulb cum spy camera (Spy Bulb) [ 16 ] , is not expected if not authorized apriori. In this scenario, we put forward three broad questions related to cyber security and management of IoT device in general and IoT cameras in particular. First, is it possible to identify the network flows of a particular IoT (spy) camera in an IoT network for better inventory management? Second, is it possible to detect an IoT camera whose flow patterns are not known apriori, i.e., zero-day IoT camera, in the presence of network flows arising from online audio/video conferencing, like Meet, Teams, and Skype, or video sharing applications, like YouTube and Prime, leading to imposing organization specific cyber security policies? Third, assuming the IoT devices can be behind NATs, is it possible to make the identification of IoT devices IP addresses and transport ports agnostic of both sources and destinations, i.e., agnostic of external systems that the devices interact with [ 17 ] . Each of these questions is related to some of the central challenges of IoT management and cyber security as per NIST [ 4 ] . To address these questions, we have overcome three main challenges. First, we have considered a large volume of network traces of 11 IoT cameras collected from our own laboratory setup and freely available real-world datasets, 4 online audio/video conferencing applications, and 2 online video-sharing applications, having overlapping cloud service providers. Second. we have carefully prepared a large volume of datasets on a case-by-case basis to consider zero-day network traffic of IoT cameras and assumed a specialized machine learning model to detect zero-day IoT cameras. Third, we have segregated the network traffic into smaller flows and extracted a large set of flow-based generic features without utilizing