The classic mental model of internet routing is a direct map: one IP address corresponds to one host, and one domain name corresponds to one distinct service. In this world, network security and traffic classification are straightforward. If you want to identify a malicious botnet, block an advertisement, or catalog the IoT devices in a household, you simply look at the source and destination identifiers in the packet headers.
Today, that simple map has collapsed. The modern internet relies heavily on shared infrastructure at both ends of the connection. At the client edge, Network Address Translation (NAT) and Virtual Private Networks (VPNs) multiplex thousands of distinct devices behind a single public IP address. At the server edge, Content Delivery Networks (CDNs) and centralized cloud platforms host millions of unrelated websites, applications, and device backends on overlapping IP pools and shared TLS certificates.
When everything looks like generic web traffic flowing between identical endpoints, traditional signature-based identification fails. A growing body of research focuses on resolving this ambiguity, developing techniques to pierce the fog of shared infrastructure by shifting focus from where traffic goes to how it behaves.
The Identity Collapse of Modern Networks
To understand why shared infrastructure thwarts traditional network analysis, consider the two primary vectors of traffic aggregation: client-side masking and server-side domain overlap.
Client-Side Masking (NATs and VPNs)
In a classic paper on spammer behavior, Y. Charlie Hu highlighted how NAT devices obscure the true volume and identity of individual actors. By aggregating an entire local network behind one public IP, an external observer cannot easily distinguish between a single high-intensity attacker and a distributed cluster of low-intensity bots.
As consumer networks have grown to encompass dozens of smart home appliances, this problem has intensified. As noted by Yair Meidan et al. in their work on A novel approach for detecting vulnerable IoT devices connected behind a home NAT, a home router collapses all internal smart home traffic into a single external identifier. Telecommunication providers see a uniform stream of bytes, making it exceptionally difficult to isolate a vulnerable IoT device from a benign laptop.
Server-Side Domain Overlap (CDNs and Cloud Backends)
On the receiving end, the rise of CDNs and cloud hosting means that destination IP addresses no longer uniquely identify a service. In Adaptive Webpage Fingerprinting from TLS Traces, Vasilios Mavroudis et al. demonstrate that webpage fingerprinting techniques relying on server IP visibility fail when traffic is routed through CDNs, as thousands of distinct sites share the exact same network identifiers.
This ambiguity is compounded by shared software infrastructure. Shahbaz Rezaei et al. point out that many mobile applications utilize identical third-party libraries for advertising and analytics. Consequently, the network flows generated by completely different apps look identical.
Similarly, Yair Meidan et al. observe that IoT devices from entirely different manufacturers frequently share the same cloud infrastructure backends (such as AWS or Google Cloud). This domain overlap creates highly similar traffic metadata profiles, leading to severe classification errors for naive monitoring tools. Malicious actors actively exploit this blind spot; for instance, Su-Chin Lin et al. discovered that advertisers deploy "replica ad domains" hosted on massive CDNs to hide their infrastructure ownership and evade ad-blocker filter lists.
[ Isolated Devices ] [ Shared Infrastructure ] [ Shared Destinations ]
Smart Fridge \ / Shared Cloud Backend
Smart Camera —-> [ Home NAT / VPN ] ===> [ CDN / Cloud ] <—————— Replica Ad Domain
User Laptop / (Single Client IP) (Single Server IP) \ Generic Web Service
Piercing the Fog with Temporal and Sequential Context
If the endpoints are anonymous and the payloads are encrypted, how can we identify what is happening inside the network? The core insight of recent literature is that shared infrastructure changes the identifiers, but it does not change the underlying behavioral mechanics of the application.
Instead of treating network packets as isolated events, researchers treat them as a continuous, contextual sequence.
Capturing Flow Sequences
While a single request to a shared analytics server is ambiguous, the context surrounding that request is highly specific. In Large-Scale Mobile App Identification Using Deep Learning, Shahbaz Rezaei et al. proposed a framework that combines a 1D CNN with an LSTM network. The CNN extracts features from individual packet bursts, while the LSTM tracks the sequence and timing of adjacent network flows. Even if an app connects to a generic, shared ad network, the specific sequence of flows preceding and following that request acts as a unique behavioral signature.
Traffic as Natural Language
This sequential perspective can be scaled up to internet-wide traffic volumes. In DeviceRadar: Online IoT Device Fingerprinting in ISPs Using Programmable Switches, Ruoyu Li et al. treat packet streams similarly to sentences in natural language processing (NLP).
By using NLP-inspired packet embeddings implemented directly on programmable hardware (P4 switches), the authors demonstrate that an Internet Service Provider can fingerprint individual IoT devices at line-rate (40 Gbps), even when those devices are multiplexed behind a NAT or encrypted inside a VPN. The structural rhythm of the device's communication—packet sizes, inter-arrival times, and protocol transitions—remains visible through the encryption layer.
Statistical Haystacks and Active Side Channels
When fine-grained packet sequences are unavailable due to sparse sampling or extreme aggregation, researchers rely on broader statistical distributions or active probing.
Sparse Metadata Analysis
At the scale of Internet Exchange Points (IXPs), inspecting every packet sequence is computationally impossible. In A Haystack Full of Needles: Scalable Detection of IoT Devices in the Wild, Said Jawad Saidi et al. proved that it is possible to detect millions of IoT devices using only sparsely sampled network flow headers.
Even when devices communicate exclusively through CDNs, blending their traffic with generic web services, the temporal distribution of their connections reveals them. Because IoT devices operate on automated, deterministic schedules (e.g., checking for updates every hour on the dot), their statistical footprint stands out against the stochastic nature of human-generated web browsing.
Exploiting Protocol Side Channels
When passive observation cannot resolve the internal hierarchy of a complex network, active side channels can force the shared infrastructure to reveal itself.
[ Off-Path Attacker ] ———( PMTU Probes )——-> [ Outermost NAT Gateway ]
|
[ Internal Chained NAT ]
|
[ Target IoT Device ]
In ReDAN: An Empirical Study on Remote DoS Attacks against NAT Networks, Xuewei Feng et al. explored how off-path attackers can map out nested, multi-layered NAT topologies. By exploiting fundamental implementation flaws in Path MTU (Maximum Transmission Unit) Discovery, the authors showed that an observer can systematically determine whether a public IP represents a single host or a complex gateway serving an internal hierarchy of chained translations.
Conclusion
Shared infrastructure was designed to optimize internet performance, conserve IPv4 addresses, and provide a baseline layer of privacy. However, it has fundamentally broken the assumption that network identifiers equal identity.
The shift toward deep-learning sequence models, NLP-style packet embeddings, and temporal metadata analysis demonstrates that traffic cannot be fully obfuscated by merely changing its routing path or sharing its IP address. By analyzing the inescapable behavioral rhythms of software and hardware, modern network measurement frameworks can successfully find the needles hidden within the internet's shared haystacks.
Key Foundational Work
Peeking into Spammer Behavior from a Unique Vantage Point (2008) — contributed 2 concept elements to this direction.
Large-Scale Mobile App Identification Using Deep Learning (2019) — contributed 1 concept elements to this direction.