ARTEMIS: Multi-Domain Research Systems

• ARTEMIS is a multi-domain label referring to distinct instruments and systems in radio astronomy, networking, and machine learning.
• In radio astronomy, it includes real-time de-dispersion systems and advanced instrumentation like ARTEMIS IV and ArTéMiS for solar and submillimeter observations.
• In networking and ML, ARTEMIS underpins BGP hijack detection, secure data processing, and efficient hardware acceleration, demonstrating broad interdisciplinary impact.

ARTEMIS is a recurrent research name rather than a single system. Across arXiv, it denotes several unrelated instruments, backends, security systems, datasets, accelerators, and learning frameworks. The best-known uses include the Advanced Radio Transient Event Monitor and Identification System for real-time radio transient searches, the Automatic and Real-Time dEtection and MItigation System for BGP hijacking, multiple astronomy instruments such as ARTEMIS IV and ArTéMiS, and a wide range of machine-learning and systems papers spanning affective vision-language modeling, homomorphic-encryption-aware pruning, zkML commit-and-prove SNARKs, in-DRAM transformer acceleration, autonomous-driving planning, and perception-policy learning (Armour et al., 2011, Sermpezis et al., 2016, Achlioptas et al., 2021, Afifi et al., 2024, Lycklama et al., 2024).

1. Names, expansions, and scope

The name is used in several technically distinct ways.

Form	Designation	Domain
ARTEMIS	Advanced Radio Transient Event Monitor and Identification System	Radio transient and pulsar backends
ARTEMIS IV	Improved multichannel solar radio spectrograph of the University of Athens	Solar radio astronomy
ArTéMiS	Wide-field submillimeter camera	Submillimeter instrumentation
ARTEMIS	Automatic and Real-Time dEtection and MItigation System	BGP hijack detection and mitigation
ArtEmis	Affective Language for Visual Art	Vision-language dataset and modeling

These usages are independent. In astronomy alone, the name may refer to a LOFAR transient backend, a solar spectrograph, a submillimeter camera, or the Artemis-enabled Stellar Imager concept for the lunar surface; in networking it identifies a BGP defense system; in machine learning and systems it labels datasets, accelerators, pruning methods, program-analysis tools, and policy-learning frameworks (Serylak et al., 2012, Kontogeorgos et al., 2010, Reveret et al., 2014, Carpenter et al., 3 Mar 2025).

2. Radio astronomy: real-time transient and pulsar processing

In high-time-resolution radio astronomy, ARTEMIS denotes the Advanced Radio Transient Event Monitor and Identification System, a combined hardware/software system for real-time processing of high-time-resolution radio astronomy data, built for searches for millisecond radio transients and pulsars with next-generation telescopes such as LOFAR and MeerKAT. The scientific target class includes pulsar single pulses, RRATs, FRBs, and related fast bursts. The central computational task is incoherent de-dispersion, implemented in a modular C++ pipeline called AMPP inside the PELICAN scalable framework developed at the Oxford e-Research Centre (Armour et al., 2011).

The pipeline receives high-time-resolution telescope streams, can perform further channelization, forms Stokes parameters, applies RFI mitigation, optionally integrates samples, then searches de-dispersed time series for single pulses or bursts. For filterbank-like total-intensity data $I(f,t)$, the cold-plasma delay is described by

$$\Delta t \approx 4.15 \times 10^3 \,\mathrm{ms}\; \mathrm{DM}\left(f_1^{-2} - f_2^{-2}\right),$$

with frequencies in MHz and $\mathrm{DM} = \int_0^d n_e(l)\,dl$. ARTEMIS uses incoherent rather than coherent de-dispersion in this work because it operates on detected intensity, is less data intensive, and maps well to GPU parallelism (Armour et al., 2011).

The 2011 system paper focuses on two brute-force de-dispersion implementations: a GPU algorithm optimized for NVIDIA Fermi and a CPU implementation optimized for Intel SSE/AVX. The GPU kernel exploits L1 cache reuse, registers for accumulators, and tiling in $(\mathrm{DM}, t)$ space; with $N_f \approx 2000$ channels and $N_{\mathrm{DM}} \approx 2000$ trial dispersion measures, it achieves about 40–50% of peak GPU performance, which the authors identify as sufficient for real-time survey operation. Simulated data tests recovered a square-pulse signal at the injected $\mathrm{DM} = 50\ \mathrm{pc\ cm^{-3}}$, and the GPU kernel was tested successfully in a real ARTEMIS environment (Armour et al., 2011).

The LOFAR deployment extended this concept into a dedicated station backend. ARTEMIS was described there as a non-disruptive add-on to LOFAR international stations, built from four 12-core servers connected through 10 Gigabit Ethernet, ingesting a 3.2 Gbit/s beamformed stream corresponding to approximately 48 MHz sky bandwidth, 5 $\mu$s sampling, and two polarizations. The pipeline performs additional polyphase-filter channelization, Stokes generation, RFI excision, temporal integration, and real-time brute-force de-dispersion over typically at least 2000 trial DMs (Serylak et al., 2012).

That paper distinguishes two GPU de-dispersion kernels. The Shared Memory algorithm has a maximum DM limit of approximately 100 and achieves ~12 Gb/s throughput on a Kepler K10 GPU. The L1 algorithm relaxes the DM limit and can process a 3.2 Gb/s single-station LOFAR stream up to DM $\approx 500$ in real time; with 10 NVIDIA K10 GPUs, it could process a 127-beam tied-array with 2000 channels per pencil beam up to DM $\approx 250$ in real time. Full systems were installed at Chilbolton and Nançay, with test systems at Jülich and Onsala (Serylak et al., 2012).

3. Astronomical instrumentation beyond the transient backend

A separate astronomical use is ARTEMIS IV, the multichannel solar radio spectrograph operated by the University of Athens at the Thermopylae Satellite Telecommunication Station. In its improved form it covers 20–650 MHz using two antennas and two receivers in parallel: a swept-frequency analyzer spanning the full band at 10 spectrums/sec with 630 channels/spectrum, and an acousto-optical receiver covering 270–450 MHz at 100 spectrums/sec with 128 channels/spectrum. The reported sensitivity is about 3 SFU in the 20–100 MHz range and 30 SFU in the 100–650 MHz range, and daily operation is fully automated, including GPS timing, antenna pointing, calibration, acquisition, and DVD archiving (Kontogeorgos et al., 2010).

Another instrument is ArTéMiS, a wide-field submillimeter camera developed for the 12-m APEX telescope. It is intended for simultaneous operation at 200 $$\Delta t \approx 4.15 \times 10^3 \,\mathrm{ms}\; \mathrm{DM}\left(f_1^{-2} - f_2^{-2}\right),$$0m, 350 $$\Delta t \approx 4.15 \times 10^3 \,\mathrm{ms}\; \mathrm{DM}\left(f_1^{-2} - f_2^{-2}\right),$$1m, and 450 $$\Delta t \approx 4.15 \times 10^3 \,\mathrm{ms}\; \mathrm{DM}\left(f_1^{-2} - f_2^{-2}\right),$$2m using filled bolometer arrays. The preliminary on-sky paper reports only the 350 $$\Delta t \approx 4.15 \times 10^3 \,\mathrm{ms}\; \mathrm{DM}\left(f_1^{-2} - f_2^{-2}\right),$$3m focal plane, with an average relative pointing accuracy of 3 arcsec, a beam estimated at 8.5 arcsec, a median NEFD at 350 $$\Delta t \approx 4.15 \times 10^3 \,\mathrm{ms}\; \mathrm{DM}\left(f_1^{-2} - f_2^{-2}\right),$$4m of 600 mJy.s$$\Delta t \approx 4.15 \times 10^3 \,\mathrm{ms}\; \mathrm{DM}\left(f_1^{-2} - f_2^{-2}\right),$$5, best values of 300 mJy.s$$\Delta t \approx 4.15 \times 10^3 \,\mathrm{ms}\; \mathrm{DM}\left(f_1^{-2} - f_2^{-2}\right),$$6, and a mapping speed already more than 5 times better than the previous 350 $$\Delta t \approx 4.15 \times 10^3 \,\mathrm{ms}\; \mathrm{DM}\left(f_1^{-2} - f_2^{-2}\right),$$7m instrument at APEX (Reveret et al., 2014).

The name also appears in the Artemis-enabled Stellar Imager (AeSI) concept, where “Artemis-enabled” refers to the NASA Artemis lunar program rather than an acronym expansion. AeSI is a proposed lunar long-baseline UV/optical imaging interferometer whose Phase I baseline design uses 15 primary mirrors arranged in an elliptical array with a 1 km major axis, with a growth path to 30 mirrors and larger arrays through staged deployments. The NIAC Phase I report argues that the lunar surface, coupled with Artemis infrastructure for transport, power, communication, and operations, makes a reconfigurable dispersed-aperture interferometer feasible (Carpenter et al., 3 Mar 2025).

4. Internet routing security: BGP hijack detection and mitigation

In networking, ARTEMIS expands to Automatic and Real-Time dEtection and MItigation System and denotes a self-operated defense against BGP prefix hijacking. Its central premise is that an AS can detect hijacks against its own prefixes with very high accuracy because it knows the legitimate origin ASNs and expected routing relationships for those prefixes. The system monitors multiple real-time control-plane sources, including BGPmon, RIPE RIS streaming, and Periscope; the 2016 full paper also discusses support for BGPstream (Sermpezis et al., 2016).

The basic detection rule is simple: if a monitored prefix is observed with an illegitimate origin AS, ARTEMIS declares a hijack. The later, more detailed work extends this to a taxonomy spanning exact-prefix hijacks, sub-prefix hijacks, BGP squatting, and Type-0, Type-1, and more general Type-N AS-path manipulations, using local knowledge of authorized origins, neighbors, and previously verified AS-links. The system is explicitly designed to be self-managed rather than third-party operated, in order to avoid delayed notification, reduce false positives, and preserve routing-policy privacy (Sermpezis et al., 2018).

Mitigation is based primarily on prefix de-aggregation. If a hijacked prefix is larger than /24, ARTEMIS can announce more specific sub-prefixes from the legitimate AS, for example splitting a hijacked /23 into two /24s. The 2017 demonstration paper, evaluated on the PEERING testbed, reports average detection in approximately 45 seconds, approximately 15 seconds to issue the de-aggregated announcements, and full mitigation within about 5 minutes, for a total of about 6 minutes from hijack launch. It explicitly notes that this method may not work for /24 prefixes because more specific IPv4 announcements are often filtered (Chaviaras et al., 2017).

The broader 2016 and 2018 studies report even faster end-to-end behavior in their own experimental settings. The 2016 paper describes detection within a few seconds and mitigation within minutes after launch, using extensive real-Internet hijacking experiments (Sermpezis et al., 2016). The 2018 paper argues that ARTEMIS combines comprehensiveness, accuracy, speed, privacy, and flexibility, and reports that prefix hijacking can be neutralized within a minute in real-world experiments (Sermpezis et al., 2018). These differing figures reflect different implementations and experimental setups rather than a single universal latency bound.

5. Machine learning, accelerators, and static analysis

In machine learning and information systems, the name covers a diverse set of unrelated contributions. ArtEmis is a large-scale vision-language dataset and modeling framework for affective language about artworks. It contains 439K emotion attributions and explanations for 81K artworks from WikiArt, supporting tasks such as emotion classification from images or text and affective caption generation. The work treats the image, the elicited emotion, and the free-form explanation as a coupled triad rather than reducing image understanding to object description alone (Achlioptas et al., 2021).

In summarization, Artemis expands to Annotation methodology for Rich, Tractable, Extractive, Multi-domain, Indicative Summarization. It is a hierarchical annotation protocol for single-document, extractive, indicative summarization, moving from paragraph-level selection to section-, document-, and short-summary selection. The paper reports analysis on 532 annotated documents, with five judges per document and average annotation time of 4.17 minutes per judge per document (Jha et al., 2020).

Several ARTEMIS papers focus on secure or efficient computation. In privacy-preserving ML under homomorphic encryption, Artemis is an HE-aware pruning method designed to reduce expensive Rotation operations in HE convolution. It uses group Lasso regularization aligned with positional and diagonal pruning patterns and reports 1.2–6x improvements for ResNet18 and ResNet50 across three datasets (Jeon et al., 2023). In hardware acceleration, ARTEMIS is a mixed analog-stochastic in-DRAM accelerator for transformer models; it combines stochastic multiplication with analog accumulation through a metal-on-metal capacitor and reports at least 3.0x speedup, 1.8x lower energy, and 1.9x better energy efficiency compared with CPU, GPU, TPU, and state-of-the-art PIM transformer accelerators (Afifi et al., 2024). In zkML, Artemis is a commit-and-prove SNARK construction compatible with any homomorphic polynomial commitment; for the VGG model it reduces commitment-check overhead from 11.5x to 1.1x (Lycklama et al., 2024).

The name also appears in program analysis and autonomous systems. A 2025 security paper presents Artemis as a static taint-analysis tool for SSRF detection in PHP web applications, combining candidate source/sink extraction, explicit and implicit call graphs, over-tainting controls, and path-condition compatibility analysis. Evaluated on 250 PHP applications, it reports 207 true vulnerable paths, including 106 true SSRFs, with 15 false positives; 35 of the detected SSRFs were newly found and reported, and 24 were confirmed and assigned CVE IDs (Ji et al., 28 Feb 2025). In autonomous driving, ARTEMIS is an end-to-end framework that combines autoregressive trajectory planning with Mixture-of-Experts, factorizing trajectory prediction as

$$\Delta t \approx 4.15 \times 10^3 \,\mathrm{ms}\; \mathrm{DM}\left(f_1^{-2} - f_2^{-2}\right),$$8

and achieving 87.0 PDMS and 83.1 EPDMS on NAVSIM with a ResNet-34 backbone (Feng et al., 28 Apr 2025).

A later multimodal-learning paper uses Artemis for structured perception-policy learning. There, the intermediate reasoning state is represented not by natural-language chain-of-thought but by proposal tuples of the form $$\Delta t \approx 4.15 \times 10^3 \,\mathrm{ms}\; \mathrm{DM}\left(f_1^{-2} - f_2^{-2}\right),$$9. Built on Qwen2.5-VL-3B, it reports strong grounding and detection performance together with generalization to counting and geometric-perception tasks, and argues that spatially grounded reasoning better matches the structure of visual perception than purely linguistic intermediate reasoning (Tang et al., 1 Dec 2025).

6. Conceptual distinctions and recurrent patterns

A common source of confusion is the assumption that ARTEMIS denotes a single evolving research program. The literature does not support that reading. The radio-astronomy backend, the BGP hijack defense system, the solar spectrograph, the submillimeter camera, the visual-art dataset, the summarization methodology, the HE pruning method, the in-DRAM accelerator, the zkML CP-SNARK, the PHP SSRF analyzer, the autonomous-driving planner, and the perception-policy learner are all separate constructions with independent authorship, objectives, and technical stacks (Armour et al., 2011, Chaviaras et al., 2017, Kontogeorgos et al., 2010, Reveret et al., 2014, Achlioptas et al., 2021, Jha et al., 2020, Jeon et al., 2023, Lycklama et al., 2024).

The spelling often signals the intended domain. ARTEMIS most often appears as an acronym in radio astronomy and networking; ArtEmis is the affective-language dataset; ArTéMiS is the APEX submillimeter camera; ARTEMIS IV is the solar spectrograph; and Artemis-enabled in AeSI refers to the NASA Artemis program. This suggests that “ARTEMIS” functions as a reused naming template across fields rather than as a stable technical brand. The repeated preference for acronymic expansion is nevertheless notable: many of these systems emphasize modularity, real-time operation, or explicit intermediate structure, whether in GPU de-dispersion, BGP mitigation, hierarchical annotation, proposal-based visual reasoning, or sequential trajectory generation (Serylak et al., 2012, Sermpezis et al., 2018, Tang et al., 1 Dec 2025, Feng et al., 28 Apr 2025).

For technical readers, the practical implication is straightforward: any reference to “ARTEMIS” requires domain disambiguation. In astronomy it may denote instrumentation or an observational backend; in Internet measurement it usually denotes a hijack-defense system; in machine learning and systems it may denote a dataset, an annotation framework, a hardware accelerator, a cryptographic compiler, or a policy-learning model. The name is therefore encyclopedically significant not because it identifies one canonical artifact, but because it recurs across multiple specialized research lineages with sharply different meanings.