{"id":"097ae6d7-2a3e-4eb3-a8d6-62a286e9b38d","arxiv_id":"2507.14842","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HiDe detects long-distance BGP interception attacks by watching per-prefix minimum round-trip times for sudden sustained jumps, using a geolocation-based lower-bound threshold, and it runs at line rate on a Tofino2 programmable switch.","lead":"This paper shows that the extra time Internet traffic takes when it is secretly rerouted through a foreign country can reveal a BGP hijack as it happens. The authors build HiDe, a programmable-switch system that detects such long-distance attacks in under a second while keeping false alarms below 0.012%.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-prefix minRTT aggregation is blind to surgical BGP hijacks that divert only a subset of flows; the §5.1 zero-false-negative guarantee does not follow from the lower-bound argument under the paper's own threat model.","rationale":"The reader's weakest assumption (geolocation accuracy) is a genuine dependency, and the paper does not stress-test it. However, I find a more load-bearing, internal problem: the per-prefix minRTT aggregation that makes HiDe scalable also nullifies the detection signal when a hijack affects only a subset of flows to the prefix. This is not a subtle data-quality caveat; it is a direct conflict with the paper's threat model, which centers on surgical interception attacks whose entire purpose is to limit the blast radius of the announcement. If the defended network is multi-homed, or if the attacker targets only one source AS, some flows to the hijacked prefix will keep using the legitimate, shorter path. Because HiDe takes the minimum RTT over all flows to the prefix, those unaffected flows pin the minRTT below tau*_mid, and the two-window detector never triggers. The §4 premise that \"all traffic to a victim prefix\" is rerouted is therefore not guaranteed under the paper's own attack model. The live experiments do not demonstrate otherwise: the Bitcoin experiment explicitly filters to European nodes before detection, and the iperf3 experiment's architecture (a transparent proxy separating the campus from the clients) makes it unclear which RTT leg is actually being measured. The trace-based evaluation in §8.1 assumes zero true positives and injects no attacks, so it provides no evidence about detection of partial hijacks. The paper's feasibility analysis (§3) also assumes all traffic between S and D is diverted, so the coverage numbers (86-97%) may be optimistic for surgical attacks. The correct resolution is not to reject the system outright, since it may work for full-prefix hijacks, but to qualify the central claim: the zero-false-negative guarantee holds only when every flow to the protected prefix is diverted, which is exactly what the stealthy surgical adversary can avoid. The authors should either restrict the threat model accordingly, change the aggregation unit to source-destination prefix pairs (and demonstrate feasibility at line rate), or empirically show that partial hijacks still raise the per-prefix minRTT in realistic multi-homed deployments. Until one of these is done, the paper's headline guarantee is unsupported. This is consistent with the reader's CONDITIONAL verdict, but the condition is more specific and more fundamental than geolocation error.","tokens_in":20360,"tokens_out":16073,"duration_ms":180840,"concrete_test":"Re-run the live iperf3/Bitcoin detection without restricting monitoring to the affected source region, and check whether the per-prefix minRTT still crosses tau*_mid when only the European flows are hijacked while other flows to the same prefix remain on the direct path. Complement with a trace replay on the campus dataset: select a destination prefix, reroute a fraction of its flows (e.g., those hashing to one upstream) onto the simulated attack path by increasing their RTT to the attack-path lower bound, and run HiDe's per-prefix minRTT changepoint detector. If no alarm fires, the zero-false-negative guarantee fails for surgical attacks.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"HiDe's detection signal is the per-prefix minimum RTT per time window (§5.2, §5.4). The zero-false-negative argument in §5.1 rests on the §4 premise that \"during a hijack, all traffic to a victim prefix must traverse the longer path via the attacker, so no RTT sample can be shorter than the minimum propagation delay via the attacker.\" But the threat model (§2.2) explicitly targets stealthy surgical interception attacks that use BGP-community manipulation (Birge-Lee et al.) to limit the blast radius of the malicious announcement. Such an attack can redirect only a subset of source networks, or only traffic via one upstream of a multi-homed defended network. Flows to the same destination prefix that are not hijacked continue to yield low RTT samples, so the per-prefix minimum stays at the benign level. Neither the absolute-threshold crossing nor the surge condition (§5.4) fires, producing a false negative for the attack class the paper claims to detect. This is an internal mismatch between the aggregated signal and the partial-flow attack model, not an external data-quality issue: it breaks the guarantee even with perfect geolocation and exact speed-of-light assumptions. The live experiments in §8.2 do not resolve this: the Bitcoin experiment filters monitoring to European nodes, and the iperf3 experiment measures a complex proxied path, so neither validates the per-prefix minRTT detector under a mixed hijacked/unhijacked flow mix. Section 8.1 injects no attacks, so it cannot test detection of partial hijacks. The geolocation concern raised by the reader is real but secondary; the subset-of-flows issue is internal and logically prior.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies detection of long-distance BGP interception attacks using propagation-delay changes. It first conducts a large-scale geometric feasibility analysis over 258 countries, reporting that even optimal (least-detectable) cross-country attacks produce RTT deviations above 5 ms for 97% of country pairs under speed-of-light assumptions and for 91%/86% of pairs under two real-measurement datasets. It then presents HiDe, a programmable-switch system that passively measures per-prefix minRTT in tumbling windows and flags a hijack when the minRTT crosses a geolocation-derived absolute lower bound tau*_mid and surges by at least lambda. HiDe is implemented in P4 on Tofino2, evaluated by faithful simulation on a 12-hour campus trace (19 billion packets) for false-positive rate and downtime, and validated with ethically conducted PEERING-based hijacks of iperf3 and Bitcoin traffic, with detection reported within 0.5 seconds.","tokens_in":20648,"tokens_out":13065,"duration_ms":159105,"significance":"If the central claims hold, this is a significant systems contribution: it shows that a data-plane-only latency signal, which the attacker cannot hide, can be turned into a line-rate detector and automatic mitigation on commodity hardware, with a clean geometric lower-bound argument and a small parameter set. The evaluation is substantial: 19B packets/7.5M flows of trace data, two real-world ethical hijacks over PEERING, Tofino2 resource measurements, and a plan to release code and anonymized latency data. The 'zero false negatives by design' claim is strong and falsifiable, which makes the correctness conditions worth scrutinizing. The paper's main limitations are that the headline coverage figures are model extrapolations rather than measured detection rates, and the threat model's surgical-attack component is not matched by the per-prefix aggregation signal.","major_comments":[{"comment":"The threat model explicitly includes stealthy surgical interception attacks that use BGP-community manipulation to limit the blast radius of the announcement (Birge-Lee et al., cited as [12]). The detection signal, however, is the minimum RTT over all flows to a destination prefix per time window (§5.2, §5.4), and the zero-false-negative argument in §5.1 relies on the premise in §4 that 'during a hijack, all traffic to a victim prefix must traverse the longer path via the attacker.' Under a surgical attack only a subset of sources, or only one upstream of a multi-homed prefix, is diverted; the remaining flows to the same prefix continue to yield low RTT samples, so the per-prefix minimum stays at the benign level and neither the absolute-threshold crossing nor the surge condition fires. This is an internal mismatch between the aggregated signal and the attack model, not a data-quality issue. The live experiments do not resolve it: the §8.2 Bitcoin experiment installs monitoring rules only for European nodes, and the iperf3 experiment measures a complex proxied path, so neither tests the per-prefix minRTT detector under a mixed hijacked/unhijacked flow mix; §8.1 injects no attacks and therefore cannot measure false negatives.","section":"§2.2, §4, §5.2, §5.4"},{"comment":"The claimed zero-false-negative guarantee rests on the auto-tuned absolute threshold tau*_mid being a true lower bound on the mid-attack RTT. This requires the geolocations of source, destination, and the threat-region boundary to be correct. Public geolocation databases (IPinfo, MaxMind) are known to be inaccurate for anycast IPs, mobile endpoints, and country-level precision; if the computed coordinates are wrong, tau*_mid can be too high, and a real attack with a shorter delay will not cross the threshold, breaking the guarantee. The paper acknowledges in §9 that free-space and satellite paths can be faster than c_f and handles them by exclusion, but geolocation error is not stress-tested anywhere in the evaluation. This is a load-bearing correctness risk for the zero-false-negative claim, not a presentation detail.","section":"§5.1"},{"comment":"The live iperf3 validation is circular with respect to the auto-tuning claim. The absolute threshold is set to 199 ms, a value estimated by traceroute during the attack ('Using multiple runs of traceroute, we estimate the lower bound of RTT as approx. 190.5 ms before attack and 199 ms during attack (absolute threshold)'). This is not the §5.1 tau*_mid computed from geolocation and speed-of-light; it is calibrated on the very attack the detector is supposed to detect. The experiment therefore demonstrates only that a threshold placed at the attack-time RTT triggers detection; it does not validate the auto-tuning component or the zero-false-negative guarantee. The Bitcoin experiment reports 'based on our calculations' without showing the threshold derivation, so it provides only weak evidence that the auto-tuned threshold is correct.","section":"§8.2"},{"comment":"The 91% and 86% coverage figures are model extrapolations, not measured detection rates. They are obtained by binning measured minOWDs by great-circle distance, fitting p25 and p75 regression lines per bin, and then substituting the p75 line for pre-attack delay and the p25 line for mid-attack delay into the geometric optimal-attack formula. No real attack is replayed on the trace, and no validation of the regression model or confidence intervals is reported. The abstract's 'at least 25%' condition is not defined in §3.5, which uses an absolute tau*_deviation >= 5 ms criterion; the two statements should be reconciled. The coverage numbers should be labeled as model-based estimates rather than measured detection rates, both in §3 and in the abstract.","section":"§3.5, Abstract"}],"minor_comments":[{"comment":"The last sentence of §5.2 says tracking minRTTs in non-overlapping tumbling windows 'requires only per-flow state,' but the system described in §6.2 maintains per-prefix state; this should be corrected to per-prefix state.","section":"§5.2"},{"comment":"The y-axis of Figure 12a begins at 190 ms, which visually magnifies a roughly 9 ms RTT increase; a full-scale plot or an explicit statement of the actual before/after deltas in the text would make the surge easier to assess.","section":"Figure 12a"},{"comment":"References [28] and [29] are the same paper ('On the Effectiveness of BGP Hijackers That Evade Public Route Collectors') and should be merged into a single entry.","section":"References"},{"comment":"The dataset-cleaning step discards flows whose minRTT indicates distances shorter than those permitted by reported geolocations; since geolocation accuracy is load-bearing for tau*_mid in §5.1, the number of discarded flows/prefixes and the effect on the resulting coverage curves should be reported.","section":"§3.5"}],"recommendation":"major_revision","confidential_remarks":"The core problem is the mismatch between the stated surgical attack model and the per-prefix aggregation signal; this should be the focus of the revision. The live validation should also be redone with a threshold derived from §5.1 rather than from attack-time traceroute values. If the authors can either restrict the threat model to whole-prefix attacks or change the aggregation signal to handle partial-flow diversion, the contribution would be much stronger."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper: the engineering is real, but the central guarantee doesn't survive contact with the threat model. HiDe detects long-distance BGP interception by tracking per-prefix minRTT and looking for a surge past a geolocation-derived absolute threshold. The authors claim zero false negatives by design. That claim rests on the statement that during a hijack all traffic to a victim prefix must traverse the longer attacker path, so no RTT sample can be below the attack lower bound. But the paper's own threat model (§2.2) explicitly targets surgical interception attacks that use BGP-community manipulation to limit blast radius. Such an attack can divert only a subset of source networks, or only traffic via one upstream of a multi-homed network. Flows to the same prefix that are not hijacked keep the per-prefix minimum at the benign level. Neither the absolute-threshold crossing nor the surge condition fires. That is an internal mismatch, not a geolocation or data-quality issue: it breaks the guarantee even with perfect geolocation and exact speed-of-light assumptions. The live experiments don't resolve it — the Bitcoin experiment filters monitoring to European nodes, and the iperf3 setup measures a proxied path rather than a mixed hijacked/unhijacked flow mix on the same prefix.\n\nNow the credit, because it's earned. The integration of per-prefix minRTT aggregation, a two-window changepoint detector, and an auto-tuned lower bound is new, and the Tofino2 prototype with low resource usage is a genuine systems contribution. The coverage analysis over 258 country pairs is interesting, the campus trace evaluation is substantial, and the ethical hijack experiments are a good-faith attempt at real-world validation. The reported sub-second detection and <0.012% FPR on the trace are plausible for full-prefix hijacks.\n\nThe soft spots beyond the load-bearing one: the absolute threshold depends on public geolocation databases, and wrong coordinates can push the threshold too high — real but secondary. The 86% coverage number is a model extrapolation using p25/p75 regression lines, not a measured detection rate. The FPR comes from one campus over 12 hours. Artifacts are promised but not yet released.\n\nThis paper deserves a serious referee, but it needs major revision. The authors should either restrict the claim to hijacks that divert all traffic to a prefix and adjust the threat model accordingly, or change the detection signal to handle partial diversion — e.g., per-source or per-flow-min tracking instead of a single per-prefix minimum. As is, the zero-false-negative guarantee is overclaimed.","headline":"A serious systems paper with a real prototype, but the zero-false-negative guarantee breaks against the paper's own surgical-attack threat model because the per-prefix minRTT signal is blind to partial hijacks.","tokens_in":21280,"tokens_out":4221,"would_cite":false,"duration_ms":51309,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A long-distance BGP hijack cannot hide: its added delay is unavoidable, and this paper builds a switch detector that flags any prefix whose minimum RTT crosses a speed-of-light lower bound, claiming zero false negatives by design.","keywords":["BGP hijack detection","BGP interception attack","propagation delay","round-trip time","changepoint detection","programmable switch","data-plane telemetry","geolocation-based threshold"],"falsifier":"Run HiDe against a real interception attack whose measured path RTT is below the auto-tuned $\\tau^*_{mid}$, for example a hijack to an anycast prefix, a path crossing a satellite or free-space link, or an attacker IP that the geolocation databases misplace by a few hundred kilometers; if the attack is not flagged, the zero-false-negative claim fails. A cheaper check is to take latency time series from past publicly documented long-distance hijacks and count how many would actually have crossed the threshold computed from today's geolocation data.","tokens_in":20113,"feed_emoji":"⚡","tokens_out":14057,"duration_ms":138053,"temperature":0.7,"pith_summary":"The paper sets out to prove that a long-distance BGP interception attack, the kind a state adversary uses to reroute domestic traffic through a foreign country for surveillance, cannot hide in the data plane, because the detour physically adds propagation delay that no attacker can conceal, and the minimum size of that delay can be computed in advance from geography. On that basis it builds HiDe, an always-on detector that watches the minimum round-trip time of each protected prefix and treats a surge past the geographic lower bound as an attack, a rule the paper states guarantees zero false negatives for the prefixes it defends. Feasibility analysis over every ordered pair of the world's 258 countries finds that even the hardest-to-detect (optimal) attacks are visible: 97% under speed-of-light idealization, and 86-91% using real-world latency measurements, with 86% of victim-attacker country pairs showing at least a 25% delay increase in production data. HiDe runs entirely on a commodity programmable switch at line rate, detected ethically conducted hijacks in under 0.5 seconds, and on a 19-billion-packet campus trace produced false positives below 0.012% with automatic rollback in a median of 0.75 seconds.","feed_headline":"Delay surge betrays long-distance BGP hijacks in half a second","feed_subtitle":"A switch-resident detector uses a speed-of-light threshold to catch detours with zero false negatives.","key_machinery":"The load-bearing object is the absolute threshold $\\tau^*_{mid}$: a lower bound on the round-trip time of any attack route, computed as the shortest great-circle path through the threat region divided by the speed of light in fiber ($c_f = 2c/3 \\approx 199.86$ km/ms), using geolocation data for the source, destination, and threat coordinates. The argument is that since a hijack reroutes all traffic to a prefix through the attacker, no RTT measured during the attack can fall below $\\tau^*_{mid}$, which makes crossing it a sufficient condition for declaring an attack. Detection itself is carried by a two-window changepoint algorithm: per-prefix RTT samples are aggregated into tumbling windows, the minimum in each window is tracked, and an attack is declared when the minRTT crosses $\\tau^*_{mid}$ between consecutive windows and rises by at least a surge threshold $\\lambda$; windows with fewer than five samples are treated as invalid. Noise is handled by min-filtering and prefix-level aggregation, and the false-positive cost is capped by an auto-correction loop that actively probes a blocked prefix and unblocks it as soon as RTT falls back below $\\tau^*_{mid}$.","core_discovery":"The central claim is that a carefully tuned changepoint detector on per-prefix minimum RTT can catch long-distance BGP interception attacks with zero false negatives by design. For each protected prefix and each user-specified threat region, HiDe computes $\\tau^*_{mid}$, the minimum possible mid-attack round-trip time, as the great-circle distance from the source to the attacker's region to the destination divided by the speed of light in fiber ($c_f = 2c/3 \\approx 199.86$ km/ms). Because a hijack moves every packet to the targeted prefix onto the attacker's path, no RTT sample observed during the attack can be smaller than this bound, so HiDe's two-window detector flags a prefix whenever the windowed minRTT crosses $\\tau^*_{mid}$ and surges by at least $\\lambda$; the paper calls the crossing an absolute proof of attack, and the guarantee against false negatives follows from the bound being a true lower bound. The claim is supported three ways: a country-pair analysis showing that optimal (least detectable) attacks still push mid-attack RTT well above pre-attack RTT for most of the world, a faithful simulation on real campus traffic showing low false positives and fast self-correction, and live ethically conducted interception attacks detected within 0.5 seconds.","pith_inferences":["The zero-false-negative claim is a property of the threshold, not of the measurement: if geolocation or the $2c/3$ fiber-speed assumption is wrong in the field, the guarantee silently degrades to a probabilistic one, so a deployment should monitor the margin between benign minRTTs and $\\tau^*_{mid}$ and treat close calls as a warning that the bound is not safe.","The same lower-bound logic extends beyond country-scale threats: any polygon-defined region yields its own $\\tau^*_{mid}$, so the scheme could defend against city- or data-center-level diversion, and operators could substitute a measured percentile of latency for the absolute bound when they prefer broader coverage over the strict guarantee.","An attacker who cannot shorten the physical detour still has one lever: inflate pre-attack minRTTs, for example by inducing congestion, to shrink the observed surge; min-filtering should blunt this, but the paper does not test an adversary who optimizes against the detector itself.","Because HiDe protects only the edge where it is deployed, it is a complement rather than a replacement for global control-plane monitoring; the two together would cover both the short-distance attacks that never cross the delay threshold and the long-distance ones that evade route monitors."],"forward_implications":["Detection time drops from the minutes that BGP control-plane convergence and route monitoring take to roughly half a second, because the surge is visible in the first packets that traverse the diverted path.","Stealth techniques that hide a hijack from BGP monitors, such as community-based surgical announcements with a small blast radius, cannot hide the delay increase, so HiDe catches attacks that control-plane systems miss by construction.","Operators get a quantitative planning tool: the country-pair analysis tells them which of their prefixes are defensible against which threat regions, and how the surge threshold $\\lambda$ trades coverage against false positives.","An always-on deployment is operationally harmless: false positives are below 0.012% on real campus traffic, and each one is corrected automatically in a median of 0.75 seconds without human intervention.","Non-TCP traffic to a protected prefix is covered for free, because any TCP flow sharing the prefix's path provides the RTT samples that reveal the diversion."],"supporting_citations":[{"why":"Supplies the physical constant $c_f = 2c/3$ for propagation through optical fiber, from which the $\\tau^*_{mid}$ lower bound is computed.","marker":"[24]"},{"why":"Dart, the line-rate RTT measurement system HiDe builds on to obtain per-flow round-trip times passively.","marker":"[36]"},{"why":"Supplies the data-ACK matching method used to compute RTTs for the 12-hour campus production trace.","marker":"[17]"},{"why":"The surgical BGP-interception technique using community manipulation that defines the threat model and that the live experiments ethically reproduce.","marker":"[12]"},{"why":"Natural Earth Admin-0 country boundary data used for all intra- and inter-country distance computations in the coverage analysis.","marker":"[18]"},{"why":"Google M-Lab NDT latency dataset providing real-world minRTT measurements behind the 86% defendability result.","marker":"[27]"},{"why":"MaxMind GeoIP, one of the geolocation services the control plane uses to convert prefixes into coordinates for the threshold.","marker":"[26]"},{"why":"IPinfo, the other geolocation service used in auto-tuning the per-prefix threshold.","marker":"[39]"},{"why":"Oscilloscope, the prior data-plane hijack detector whose false-positive and false-negative behavior HiDe is positioned against.","marker":"[14]"},{"why":"PEERING testbed, which supplies the announceable prefixes and distributed vantage points enabling the ethical live hijack validation.","marker":"[35]"}],"fun_headline_variants":["Latency spike flags BGP hijacks in under half a second","Data-plane latency catches cross-border BGP hijacks","Zero false negatives for long-distance hijacks via RTT","Speed-of-light bound exposes BGP reroutes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The zero-false-negative guarantee holds only if $\\tau^*_{mid}$ is a true lower bound on the delay of every real attack route, which assumes the geolocation databases locate the source, destination, and threat region correctly and that no path is faster than two-thirds of the speed of light along the great-circle distance; a wrong location (anycast prefixes, country-level precision, satellite-backed paths) would set the threshold too high and let a genuine long-distance attack pass unflagged.","fun_headline_variants_meta":{"raw":{"variants":["Latency spike flags BGP hijacks in under half a second","Data-plane latency catches cross-border BGP hijacks","Zero false negatives for long-distance hijacks via RTT","Speed-of-light bound exposes BGP reroutes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000498,"raw_usage":{"total_tokens":2497,"prompt_tokens":1061,"completion_tokens":1436,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":677,"completion_tokens_details":{"reasoning_tokens":1370}},"tokens_in":677,"tokens_out":1436,"duration_ms":12548,"temperature":1.0,"reasoning_tokens":1370,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:46:40.482797+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run HiDe against a real interception attack whose measured path RTT is below the auto-tuned $\\tau^*_{mid}$, for example a hijack to an anycast prefix, a path crossing a satellite or free-space link, or an attacker IP that the geolocation databases misplace by a few hundred kilometers; if the attack is not flagged, the zero-false-negative claim fails. A cheaper check is to take latency time series from past publicly documented long-distance hijacks and count how many would actually have crossed the threshold computed from today's geolocation data.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the physical constant $c_f = 2c/3$ for propagation through optical fiber, from which the $\\tau^*_{mid}$ lower bound is computed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Dart, the line-rate RTT measurement system HiDe builds on to obtain per-flow round-trip times passively."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the data-ACK matching method used to compute RTTs for the 12-hour campus production trace."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The surgical BGP-interception technique using community manipulation that defines the threat model and that the live experiments ethically reproduce."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Natural Earth Admin-0 country boundary data used for all intra- and inter-country distance computations in the coverage analysis."},{"cited_title":"(2024-12-01 – 2024-12-05)","cited_arxiv_id":null,"evidence_quote":"Google M-Lab NDT latency dataset providing real-world minRTT measurements behind the 86% defendability result."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MaxMind GeoIP, one of the geolocation services the control plane uses to convert prefixes into coordinates for the threshold."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"IPinfo, the other geolocation service used in auto-tuning the per-prefix threshold."},{"cited_title":"Oscilloscope: Detecting BGP Hijacks in the Data Plane","cited_arxiv_id":"2301.12843","evidence_quote":"Oscilloscope, the prior data-plane hijack detector whose false-positive and false-negative behavior HiDe is positioned against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PEERING testbed, which supplies the announceable prefixes and distributed vantage points enabling the ethical live hijack validation."}],"review_version":1}