Deanonymizing Device Identities via Side-channel Attacks in Exclusive-use IoTs & Mitigation

Deanonymizing Device Identities via Side-channel Attacks in Exclusive-use IoTs & Mitigation

Speakers: Christopher Ellis (PhD Student, Ohio State University)

Conference: NDSS Symposium

YouTube: https://www.youtube.com/watch?v=WmBPdX0IWEM

Overview

This talk, presented by Christopher Ellis, a PhD student at Ohio State University, unveils a critical and historically overlooked flaw in the communication patterns of exclusive-use IoT devices. Titled "Deanonymizing Device Identities via Side-channel Attacks in Exclusive-use IoTs & Mitigation," the research introduces a new class of tracking attacks, dubbed ID bleed, which exploit a fundamental boolean indicator of trusted relationships. This vulnerability allows adversaries to deanonymize and track devices even when modern countermeasures like MAC address randomization are in place.

The research not only exposes this profound weakness across ubiquitous protocols like Bluetooth Low Energy (BLE) and Wi-Fi but also proposes a practical and efficient mitigation framework called the anonymization layer. This defense mechanism is designed to remove the observable side channel with minimal overhead, demonstrating a path forward for securing device identities. The findings are significant for IoT manufacturers, protocol designers, and anyone concerned with the privacy and security of connected devices, highlighting that even seemingly secure communication can leak sensitive identifying information.

Background

▶ Watch: Introduction to ID bleed tracking attacks (0:00)

The pervasive nature of wireless communication, particularly in the realm of Internet of Things (IoT) devices, presents a persistent challenge: packets are inherently observable. While this observability is a foundational aspect of network operation, it also creates a powerful capability for adversaries. The researchers embarked on an investigation into network traffic captures from various IoT devices, seeking to identify any extractable information or insightful patterns, regardless of whether devices were actively paired or not.

What they uncovered was a subtle yet profound side channel effect: a boolean indicator reflecting the trusted relationship between devices. This effect manifests in "exclusive-use devices," defined as systems with a one-to-one or one-to-many pre-established trusted relationship. The essence of this side channel is simple: when a trusted device (e.g., Alice) makes a request to another device (e.g., Bob), Bob responds. However, when an untrusted device (e.g., Charlie) makes a similar request, Bob either does nothing or provides a distinctly different, often absent, response. This difference in response behavior—the presence or absence of a valid reply—serves as an observable signal that allows an adversary to infer the existence of a trusted relationship.

This concept, while seemingly straightforward, becomes significantly more potent because these ID bleed attacks are capable of defeating modern privacy countermeasures like MAC address randomization. Previously, rotating MAC addresses was considered a robust method to prevent long-term tracking. However, the ID bleed attacks demonstrate that even with ephemeral identifiers, the underlying communication logic can still betray device identity and location. The research highlights that while the idea of inferring information from communication patterns has been discussed, its implications for tracking users in the hands of an adversary have been "historically overlooked."

Key Findings

▶ Watch: ID bleed defeats MAC address randomization (2:45)

The research presents two primary categories of ID bleed attacks: passive and active, both capable of deanonymizing exclusive-use IoT devices and enabling tracking.

1. Passive ID Bleed Attacks:

These attacks rely solely on observing communication patterns over time. Even when devices employ MAC address randomization and rotate their ephemeral identifiers (e.g., A1 to A2 for Alice, C1 to C2 for Charlie), the consistent presence of a response from the trusted device (Bob) to Alice's varying identifiers, coupled with the absence of a response to Charlie's identifiers, reveals the underlying trusted relationship. An adversary can simply monitor the airwaves and deduce which ephemeral IDs belong to a trusted device based on the responses they elicit. This allows for linkage of different ephemeral identifiers to the same trusted entity over time.

2. Active ID Bleed Attacks:

More powerful and versatile, active attacks enable an adversary (Eve) to actively probe devices and track their location. In this scenario, Eve can capture an unknown identifier (e.g., an ephemeral MAC address) from a device at a specific location. Eve then relays this identifier to the target device (Bob), which might be in a different location. If Bob responds, Eve infers that the originating device (Alice) was indeed at the location where the identifier was captured. Conversely, if Eve relays another identifier and Bob does not respond, Eve can infer that the device is not present at that specific location. Crucially, Eve only needs to observe Bob's response; relaying the response back to the original device is not necessary for tracking, though it could be used for protocol manipulation. This method allows for precise location tracking of devices across various points in time and space.

Vulnerable Protocols and Applications:

The researchers specifically investigated common and ubiquitous IoT protocols and applications, finding them susceptible to both passive and active ID bleed attacks:

• Bluetooth Low Energy (BLE): Vulnerabilities were found in features related to confidentiality, secure connections, connection gating, signing procedures, authentication, and autoconnection. For instance, simply observing an encryption request followed by an encryption response (passive) or relaying packets during the encryption negotiation stage and observing the response (active) was sufficient.
• Replay Attacks in BLE: The BLE encryption negotiation stage was also found to be vulnerable to replay attacks. An adversary can capture legitimate packets from a trusted smartphone, replay them to the peripheral at a later time, and observe a valid encryption response. If packets from an unknown smartphone are replayed, a rejection message would be observed, further confirming the side channel.
• Wi-Fi: Similar vulnerabilities were identified in Wi-Fi, particularly during the authentication stage, even before standard probe requests and responses.
• IoT Smartphone Companion Apps: The research extended to real-world IoT companion applications controlling devices like smart lights, blood pressure monitors, smart locks, and smart plugs. These apps were found to leak the boolean indicator during authentication attempts. For example, even an incorrect password attempt when trying to authenticate with smart lights would still leak this crucial information, indicating a trusted relationship (or lack thereof) through the presence or absence of a specific response.

Distributed Relay and Replay Attack Network:

To illustrate the potential scale of these attacks, the talk outlined a conceptual architecture for a distributed attack network. In this setup, an adversary could deploy multiple "Eve" nodes across various locations (e.g., an office, a person's home). These nodes would capture packets from a target peripheral (Bob), transmit them to a central attack network server, which could then broadcast or publish these identifiers to any number of other deployed nodes. This distributed network would enable widespread deanonymization and tracking of devices across multiple physical locations, highlighting the severe privacy implications of ID bleed.

Technical Deep Dive

▶ Watch: How the passive ID bleed attack works (3:00)

The core technical insight of this research lies in identifying and exploiting the boolean indicator of trust in exclusive-use IoT devices. An exclusive-use device is defined as any device or system that maintains a one-to-one or one-to-many previously trusted relationship with another entity. This trust is typically established during a pairing or authentication phase, and subsequent interactions are governed by this pre-existing bond.

The side channel emerges from the differential behavior of the trusted device (Bob) when interacting with a trusted requestor (Alice) versus an untrusted one (Charlie). If Alice sends a valid request, Bob processes it and sends a response. If Charlie sends a similar request but is untrusted, Bob either provides no response or a response that is distinctly different, often a rejection or error message. The presence or absence of a valid, expected response is the observable side channel that ID bleed attacks leverage.

Passive Attack Mechanism:

Consider Alice and Charlie, both attempting to communicate with Bob. At time t1, Alice uses a random address A1, and Charlie uses C1. Alice makes a request, Bob responds. Charlie makes a request, Bob does nothing. At time t2, Alice and Charlie rotate their ephemeral identifiers to A2 and C2 respectively. Again, Alice's request (from A2) elicits a response from Bob, while Charlie's request (from C2) does not. By observing this consistent pattern of responses to certain ephemeral identifiers (A1, A2) and lack of responses to others (C1, C2), an adversary can infer that A1 and A2 belong to the same trusted entity (Alice), effectively linking her randomized MAC addresses and tracking her over time. This works because the underlying protocol logic still differentiates between trusted and untrusted requests, regardless of the ephemeral identifier used.

Active Attack Mechanism:

The active attack amplifies the passive method by allowing an adversary (Eve) to actively probe devices for responses.

1. Capture and Relay: Eve, positioned at location 'A', captures an unknown ephemeral identifier from a device (say, Alice).
2. Probe the Target: Eve then relays this captured identifier, along with a crafted request, to the target device (Bob). Bob could be in the same location or a different one.
3. Observe Response: Eve observes Bob's behavior. If Bob responds with a valid message, Eve infers that the captured identifier belongs to a device trusted by Bob, and thus Alice was present at location 'A' at the time of capture.
4. Negative Inference: To confirm Alice's absence from a location, Eve can capture another ephemeral identifier at location 'A' (e.g., from Charlie, or from Alice at a later time if she has moved). If Eve relays this new identifier to Bob and Bob does not respond, Eve can infer that the trusted device (Alice) is no longer at location 'A' or that the new identifier belongs to an untrusted device. This allows for precise, real-time tracking of a device's presence or absence at specific locations. The key is that Eve only needs to observe Bob's response, not necessarily complete a full communication handshake.

Vulnerability Examples:

• BLE Encryption Negotiation: During the BLE encryption negotiation stage, a trusted device sends an encryption request and receives an encryption response. An active attacker can capture this request, replay it to the peripheral, and observe the response. If the peripheral responds, the attacker knows the replayed packet originates from a trusted device. If an untrusted packet is replayed, a rejection message is observed, creating the side channel. This is a powerful replay attack vector.
• Wi-Fi Authentication: The authentication phase in Wi-Fi, prior to probe requests and responses, also exhibits this differential behavior. A device attempting to authenticate with a known access point will elicit a different response pattern than one attempting to authenticate with an unknown or incorrect credential.
• IoT Companion Apps: Even high-level application-layer authentication in smartphone companion apps for IoT devices (e.g., a smart lock, smart lights) leaks this boolean indicator. An attempt to authenticate with an incorrect password might still trigger a specific (even if negative) response from the device, whereas an utterly unknown request might be ignored, creating the side channel.

Anonymization Layer (Mitigation):

To counter ID bleed, the researchers propose the anonymization layer, a mitigation framework designed with several key features:

1. Remove Transmission Direction (Unicast to Broadcast): Instead of one-to-one unicast communication, the layer shifts to a broadcast model. This prevents data flow analysis that relies on source-destination relationships.
2. Hide Context (Encryption): All packet types, metadata, and context are encrypted. This ensures that an adversary cannot infer information from the structure or content of packets.
3. Provide Entropy: Consistent payloads are randomized or obfuscated so they don't appear as identical encrypted bytes, preventing pattern analysis.
4. Tunable Pseudo-Responses: Crucially, the anonymization layer introduces responses for untrusted devices that are indistinguishable from legitimate traffic. This removes the "no response" or "different response" side channel, creating ambiguity for the adversary.
5. No Communication Stack Modification: Designed as a separate, additional layer, it can be integrated without modifying the underlying communication stack. This can be achieved using mechanisms like eBPF or kernel modules for deployment.

Anonymization Layer Key Exchange and Packet Structure:

The process involves:

• Generating and exchanging random keys between Alice and Bob.
• XORing these keys to create a pairing key.
• Using the pairing key to generate source keys and the base of encryption keys.
• Generating a set of ephemeral transmission keys (which act as source identifiers) and corresponding paired encryption keys for each transmission key.
• The final anonymization layer packet consists of a single source identifier (from the ephemeral set) and an encrypted payload. This ensures that all packets appear uniform and unlinkable from an external observer's perspective.

Performance Evaluation:

The anonymization layer was evaluated for its overhead:

• Throughput Overhead: On both PC and smartphone platforms, the lookup and encryption processes introduced an overhead of under 2% for throughput, demonstrating high efficiency.
• Key Resolution Overhead: Two methods were evaluated:
• Hash Method: The existing method used in BLE for IRK (Identity Resolving Key) resolution. For a smartphone with 16 paired devices, this resulted in approximately 7% overhead. On a PC (representative of a server), it was 55%.
• Cache Method: The proposed new method significantly outperformed the hash method. For a smartphone with 16 pairs, overhead was sub-1%, and on a PC, it was just over 1%. This highlights the efficiency of the cache-based key resolution.
• Power Overhead: Measured over 60 minutes, with varying frequencies (0%, 50%, 100%) of pseudo-responses, the power consumption overhead was consistently less than 2%.

These performance metrics underscore the practicality of the anonymization layer as a viable mitigation strategy that can be adopted without severely impacting device performance or battery life.

Demo / Proof of Concept

▶ Watch: Real-world IoT protocols vulnerable to ID bleed (4:35)

While the talk describes the mechanisms of the ID bleed attacks and the anonymization layer in detail, it primarily focuses on the conceptual framework, implementation specifics, and performance evaluation rather than a live demonstration of a specific proof-of-concept during the presentation. The comprehensive analysis of vulnerabilities across BLE, Wi-Fi, and IoT companion apps, along with the detailed design and performance evaluation of the anonymization layer, serves as a robust theoretical and empirical proof of concept. The described "distributed relay and replay attack network" acts as an architectural blueprint for how such attacks could be practically deployed on a larger scale, illustrating the feasibility and impact of the identified vulnerabilities. The source code for the anonymization layer is also stated to be made available, which would allow others to replicate and test the proposed mitigation.

Defensive Implications

▶ Watch: Active attack enables replay attacks (6:00)

The findings from this research have profound implications for various stakeholders in the IoT ecosystem, necessitating a fundamental shift in how device identity and privacy are approached.

For IoT Device Manufacturers and Protocol Designers:

• Re-evaluate Communication Patterns: The most critical takeaway is to meticulously scrutinize all communication stages for any differential behavior that could act as a boolean indicator of trust. This includes authentication, pairing, encryption negotiation, and even simple data exchange. The "doing nothing" response to an untrusted request is a critical vulnerability.
• Implement Indistinguishable Responses: Adopt mechanisms like the proposed anonymization layer to ensure that responses (or lack thereof) to both trusted and untrusted requests are indistinguishable to an external observer. This involves providing tunable pseudo-responses for untrusted requests that mimic legitimate traffic.
• Move Beyond MAC Address Randomization: MAC address randomization, while a valuable privacy feature, is insufficient on its own. It only obfuscates the link-layer identifier; the underlying protocol logic can still leak identity. Holistic privacy-by-design principles must be applied at all layers of the communication stack.
• Address Replay Attacks: Protocols involving encryption negotiation or authentication that are vulnerable to replay attacks (as seen in BLE) need to incorporate robust anti-replay mechanisms that are resilient to observed side channels.
• Consider Layered Defenses: The anonymization layer is designed as a separate, additional layer, making it easier to integrate into existing systems without requiring full communication stack overhauls. This approach should be considered for rapid deployment.

For Network Defenders and Security Researchers:

• Awareness of Side Channels: Understand that network traffic analysis can reveal more than just explicit data; implicit behaviors (like response patterns) can be equally revealing.
• Monitoring for ID Bleed: While challenging, defenders might look for patterns of communication that exhibit the characteristics of ID bleed attacks, particularly active probing attempts followed by differential responses.
• Adoption of Anonymization Layer: Advocate for and implement the anonymization layer or similar privacy-enhancing technologies in their IoT deployments, especially for devices handling sensitive information or operating in environments where user tracking is a concern.

For End-Users:

• Increased Privacy Awareness: Users should be aware that even seemingly secure IoT devices, including those with companion apps, can potentially leak their identity and location through these side channels.
• Demand Secure Products: Support manufacturers who prioritize robust privacy and security features that go beyond basic encryption and MAC address randomization.

In essence, the research mandates a shift from merely securing the content of communication to securing the patterns of communication. The presence of a fundamental flaw in communication patterns means that protocols, irrespective of their specific characteristics (BLE, Wi-Fi, ZigBee, Z-Wave), are susceptible if they exhibit this boolean difference in response to trusted versus untrusted entities. The proposed anonymization layer offers a practical and efficient blueprint to address this pervasive vulnerability.

Key Takeaways

• Fundamental Flaw: Exclusive-use IoT devices exhibit a fundamental and historically overlooked flaw in their communication patterns, creating a side channel based on differential responses to trusted versus untrusted requests.
• ID Bleed Attacks: These novel attacks exploit this side channel to deanonymize device identities and enable user tracking, even when modern privacy countermeasures like MAC address randomization are employed.
• Passive and Active Vectors: Both passive observation of communication patterns and more powerful active relay/replay attacks are effective across ubiquitous IoT protocols (BLE, Wi-Fi) and smartphone companion apps.
• Broad Vulnerability Scope: Specific vulnerabilities were identified in BLE (confidentiality, authentication, encryption negotiation, replay attacks), Wi-Fi (authentication), and companion apps for smart lights, blood pressure monitors, smart locks, and smart plugs.
• Anonymization Layer Mitigation: A practical and efficient mitigation framework, the anonymization layer, is proposed. It removes the side channel by shifting to broadcast communication, encrypting context, providing entropy, and crucially, offering indistinguishable pseudo-responses for untrusted requests.
• Low Overhead Defense: The anonymization layer demonstrates minimal performance impact, with less than 2% throughput and power overhead, and highly efficient key resolution using a novel cache method (sub-1% for smartphones).

About the Speaker(s)

The primary speaker for this presentation was Christopher Ellis, a PhD Student at Ohio State University. He introduced himself as a first-time NDSS presenter and an NDSS Internet Society fellow, highlighting his active involvement in the security research community. Christopher is also a vulnerability researcher and a member of the computer security laboratory at Ohio State University. His work on this research, focusing on deanonymizing device identities, was conducted in collaboration with his colleagues UA Jang, Mojit Kumar Jangid, Shu Shen Jao, and his advisor Dr. Gi Changeng Lin.