Decentralized storage networks such as IPFS have a structural privacy flaw: to find a file, a node broadcasts what it wants — a content identifier, or CID — to every peer it is connected to. Balduf et al. ran passive monitor nodes for 15 months, recorded 27.8 billion of these request messages, and showed that anyone patient enough can log who asks for what, recovering the mapping between a user (the "principal") and their network address. They even identified 93 gateway nodes belonging to major operators with what they call "negligible deniability," simply by injecting unique CIDs through public web gateways and watching which node then broadcast a want for them. The obvious fix — flooding the network with decoy requests so the real one disappears into noise — fails on the numbers: over 80% of CIDs in their trace were requested by only a single peer, so there is no natural crowd to hide in, and naive decoys multiply bandwidth across peers that can ill afford it.
Rare-event physics solved a structurally identical problem decades ago. Experiments searching for dark matter or neutrinoless double-beta decay face one true signal event buried under millions of background events, and they separate them not by brute-force shielding alone but by morphology: nuclear recoils in liquid argon produce a sharp burst of prompt light while electronic-recoil backgrounds produce slow, smeared light, letting DEAP-3600 reach an electronic-recoil leakage below $1.2 \times 10^{-7}$ at 90% confidence in the 15–31 keVee range (Amaudruz et al.); in germanium detectors, surface alpha backgrounds betray themselves in the slow tail of the waveform, which the Delayed Charge Recovery technique uses to reject 100% of surface alphas while sacrificing only 0.2% of bulk events (Arnquist et al.). The crucial property of these discriminators is asymmetry: the detector, holding the calibration and the raw waveform, can compute the pulse-shape statistic; a cruder observer cannot. This proposal imports that asymmetry into peer-to-peer request privacy.
Core Claim
A client can bundle each real want-broadcast into a small burst of decoy wants whose content, timing, and follow-up behavior are statistically indistinguishable from ambient network background, while a keyed pulse shape — a micro-timing watermark derived from a shared secret with the intended provider — lets that provider alone recover the real request with near-zero loss. To a passive monitor without the key, the members of the burst are exchangeable, so its posterior on which CID the principal actually wants collapses to chance. Formally, for a burst $B$ of $d{+}1$ requests (one real, $d$ decoys) and monitor observations $\mathcal{O}$ visible without the key:
$$P(\text{principal wants } x_i \mid \mathcal{O},\ \text{no key}) = \frac{1}{d+1} \quad \text{for every } x_i \in B,$$
which is the networking analogue of an anonymity set, and the quantity we will measure exactly the way Veske et al. measure astrophysical confidence: their exact Bayesian framework lifts subthreshold gravitational-wave candidates from $p_{\text{astro}} \sim 0.1$ to $\sim 0.9$ by combining temporal, spatial, and energy likelihoods across observatories. We invert that machinery into an adversary model — treating the monitor's timing, content-popularity, and response-traffic observations as independent "messengers" whose joint likelihood determines the deanonymization posterior — and then design the defense so that posterior stays at chance. Direction two of this proposal therefore plays two roles: pulse-shape discrimination supplies the defensive mechanism, and the Bayesian coincidence calculus supplies the offensive yardstick.
Research Program
The work has three phases. First, we build a calibrated IPFS request simulator from the two measurement studies: the 15-month passive trace of Balduf et al. (roughly 10,000–14,000 active nodes, 70% of activity in the US, Netherlands, and Germany, the single-peer CID popularity skew) and the crawler data of Henningsen et al., whose churn numbers — 44% of node sessions under five minutes, 52% of nodes behind NATs — are essential because any burst-based defense must survive peers disappearing mid-exchange. Second, we design the keyed pulse-shape channel: the client emits the burst with inter-packet delays drawn from a key-seeded schedule, decoy CIDs sampled from the popularity distribution calibrated on the real trace (this is the background-matching step every rare-event search performs before cutting anything), and the provider computes a discriminator statistic on the timing trace — the direct analogue of the prompt-to-total light ratio $F_{\text{prompt}}$ in DEAP-3600 — to pick out the real want. Third, and most honestly difficult, we handle the response channel: a monitor that is itself a connected peer sees which want leads to a block transfer, so the provider must make its responses background-like as well, for instance by serving matched-size junk blocks for decoys. The core quantitative deliverable is the privacy-per-decoy-byte curve: monitor attribution posterior versus bandwidth overhead, for keyed bursts versus uniform cover, across burst sizes and churn rates.
This is deliberately different from the existing work at this boundary. Loopix also injects background noise (loops and drop cover traffic), but it is unkeyed, uncoordinated Poisson traffic for store-and-forward messaging with end-to-end latencies in the order of seconds and relay throughput over 300 messages per second — a regime incompatible with block retrieval, and, more fundamentally, its noise carries no morphology an authorized receiver can discriminate. Balduf et al. themselves suggest "cover traffic or salted CID hashes" but provide no framework, protocol, or evaluation; we supply both, and we show why salting alone is orthogonal (it hides what is asked, not who asked). Backes et al. hide DHT lookup keys with oblivious transfer at only a few milliseconds of overhead, but their scheme is strictly cryptographic, assumes structured quorum routing, and explicitly leaves open range-estimation attacks — it does not touch unstructured broadcast protocols like BitSwap at all. APAC replaces direct links with anonymous circuits, achieving an anonymity score of 0.8 with two relays under 35% malicious nodes, but circuits assume an honest-but-curious adversary and are torn apart by the churn documented by Henningsen et al.; our defense applies precisely where circuits cannot — promiscuous broadcast.
Baselines
The "do nothing" reference point is the deployed BitSwap protocol as measured by Balduf et al.: over 15 months their monitors identified 93 gateway node IDs with negligible deniability, and their trace statistics define the background any defense must blend into. We will replay their gateway-deanonymization attack verbatim in the simulator — inject unique probe CIDs through gateway interfaces, watch broadcast propagation — and any defense must be judged against this attack first; a defense that leaves probe-CID attribution above chance has failed regardless of its other numbers.
The strongest existing competitor is uniform Poisson cover traffic in the style of Loopix's drop cover, adapted to want-broadcasts: the client emits decoy wants as a Poisson process at rate $\lambda$ with CIDs drawn from the popularity distribution, and the provider responds to everything it can. This is the mechanism the mix-network literature actually validated (Piotrowska et al.), and it requires no key distribution, no clock synchronization, and no protocol change beyond message volume — which is exactly why a reviewer would prefer it. No reported result found on Lacuna for uniform cover traffic applied to BitSwap-style broadcast retrieval, so we will run it ourselves in the same simulator, on the same trace-calibrated background, with the same bandwidth budget as the keyed scheme (matched decoy-byte overhead, swept over ratios from 2× to 20×), and tune its rate parameter with the same search budget we give our own burst size and key schedule. Beating it means: at every matched overhead, keyed pulse shaping keeps the monitor's Bayesian attribution posterior at chance ($1/(d{+}1)$ per burst member) while uniform cover leaves it significantly above chance — because the real want's response traffic and timing stand out — or, if uniform cover closes that gap by covering responses too, keyed shaping must achieve the same privacy at a strictly lower byte cost. If neither holds, the keying adds nothing and the proposal is wrong.
Two further baselines sharpen the comparison. Salted CID hashing, as suggested by Balduf et al., will be implemented and run to demonstrate its orthogonality: it should leave principal-to-address attribution untouched while hiding content identity. And the key ablation — identical bursts with the key removed, so the provider discriminates by watching which want completes a transfer — isolates the proposal's central ingredient; the monitor should then succeed, since the discriminating signal is exactly the observable the monitor can also read.
Why not just…?
Why not just route everything through Tor or anonymous circuits? Jia et al. state directly that Tor's multi-hop overhead is unsuitable for low-latency web content delivery, which is why they built APAC's WebRTC circuits; but APAC assumes honest-but-curious peers and breaks under Sybil attacks, and the churn measured by Henningsen et al. — 44% of sessions under five minutes — collapses circuits faster than they can be rebuilt. More basically, circuits protect point-to-point connections; BitSwap's leak is a broadcast to thousands of connected peers, which circuits do not wrap at all. Our defense targets the broadcast itself.
Why not just salt or encrypt the CIDs? Salting hides which content is requested but not which node requested it — the monitor still learns that this address broadcast a want, and the principal-address mapping survives intact, which is precisely the mapping Balduf et al. exploited to unmask gateways. It also forces every provider to brute-force salts for every incoming want. We will run salted hashing as a baseline to make this orthogonality empirical rather than rhetorical.
Why not just inject uniform Poisson cover traffic? This is the real threat to the proposal, and we do not argue around it — we measure it. The mix-network evidence for uniform cover comes from store-and-forward messaging with second-scale latencies (Piotrowska et al.), not from high-throughput block retrieval, and nobody has reported results for it at BitSwap scale (no reported result found on Lacuna for that setting). The honest possibility is that uniform cover with response covering matches our privacy at equal cost; the privacy-per-decoy-byte experiment, run at matched overhead with matched tuning, is designed precisely to settle this, and the key ablation tells us whether any observed advantage actually comes from the keyed pulse shape rather than from the burst structure or popularity matching. If uniform cover wins, that is the paper's finding.
Why not just tell users to use a VPN or OS privacy extensions? The measurement literature rules this out at the boundaries: Nikiforakis et al. showed browser plugins bypass host proxies entirely, Nibert et al. showed WebRTC's path discovery leaks true public and private IPs straight through active VPNs, and Saidi et al. showed one legacy EUI-64 device in a prefix defeats ISP rotation for the whole household, enabling tracking of up to 17% of subscriber lines. Endpoint tools cannot fix a protocol that broadcasts interests by design.
Proposed Experiments
The decisive experiments, in order. First, attack replay: reproduce the Balduf gateway attack on the trace-calibrated simulator and confirm the do-nothing baseline achieves near-certain gateway identification, matching the qualitative "negligible deniability" finding. Second, the privacy-per-byte sweep: for burst sizes $d$ from 1 to 32 and matched uniform-cover rates, plot the monitor's Bayesian attribution posterior and the fraction of probe CIDs correctly attributed; the pass condition is posterior at chance for keyed bursts at every overhead where uniform cover remains above chance, or equal privacy at strictly lower cost. Third, the learned adversary: train a strong classifier on everything a keyless monitor sees — timing traces, CID popularity, response correlations — and require area under the ROC curve statistically indistinguishable from 0.5 against keyed bursts, while confirming the same classifier demolishes the unkeyed ablation. Fourth, signal efficiency: the provider must recover real wants with at least 99% efficiency and near-zero false serves, the networking counterpart of the PandaX-III tracking cuts, which suppressed gamma backgrounds by roughly seven orders of magnitude while keeping signal efficiency above 20% (PandaX-III Collaboration); we hold ourselves to the stricter DEAP standard of leakage below $10^{-7}$ as the aspirational ceiling for decoy-as-real confusion. Fifth, churn robustness: with 44% of sessions dying under five minutes, measure how often bursts are orphaned and whether the keyed recovery degrades gracefully. The whole program is simulation-first on published traces, so it is cheap, falsifiable, and — if the privacy-per-byte curve comes out the wrong way — decisive in the useful direction.
References
- Monitoring Data Requests in Decentralized Data Storage Systems: A Case Study of IPFS
- First Results from the DEAP-3600 Dark Matter Search with Argon at SNOLAB
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- Enhancing the significance of astrophysical events with multimessenger coincidences
- Mapping the Interplanetary Filesystem
- The Loopix Anonymity System
- Adding query privacy to robust DHTs
- Anonymity in Peer-assisted CDNs: Inference Attacks and Mitigation
- Cookieless Monster: Exploring the Ecosystem of Web-Based Device Fingerprinting
- Preventing WebRTC IP Address Leaks
- One bad apple can spoil your IPv6 privacy
- Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment