\section{Evaluation protocols and evaluability} \label{sec:protocols} \subsection{Population, feature sets and leak flags} The evaluation population contains the first occurrence of every pair of feature vector and label on feature set A (64,475,595 flows). Deduplicating on the vector alone would silently delete one label from every label-conflicting group, which is why the label is part of the key. We use two feature sets: A with all 53 non-label features, and B with the 47 features that remain after dropping IP addresses, ports and timestamps. Many B vectors recur between training and test sets of the same protocol, so each test flow is flagged when its B vector occurs in the protocol's training set with the same label (a duplicate that makes the test easy) or only with other labels (an irreducible label conflict). Scores for feature set B are computed on test flows that carry neither flag; scores for A use all test flows of the population. Because the evaluable label set can differ between A and B, scores for A and B refer to different label sets and test rows, and we report a matched-label sensitivity analysis where this matters (\Secref{sec:results}). \subsection{Protocols} A label counts as evaluable under a protocol if at least 100 flows are available for training and at least 50 for testing; for feature set B the training count is the number of distinct training vectors and the test count is the number of clean test flows. Only evaluable labels are trained and scored, and the evaluability matrix lists them. The protocols are summarised in \Tabref{tab:protocols}. \begin{description} \item[P1, random (reference).] Flows are split 70/15/15 at random, stratified within each pair of source and class. Within-source random splits are common in the literature; we use P1 only as a reference because it shares the most structure between training and test sets. \item[P2, per-class chronological.] Within every pair of source and class the flows are ordered by start time; the first 70\% form the training set, the next 15\% the validation set and the last 15\% the test set. Flows within an embargo of five minutes (or 1\% of the stream's duration if shorter) around each boundary are dropped so that concurrent flows of one session do not straddle a boundary. Attacks occupy one to three capture days, so P2 tests generalisation within an attack episode, not across episodes. The windows differ per class; with feature set A the absolute timestamp can therefore separate classes, and headline P2 scores use feature set B. \item[P3, source-IP groups.] Within every pair of source and class, all flows of a source IP are assigned to the same split (70/15/15 by a hash of the address). A label counts as evaluable only with at least five training and five test source IPs and a top-IP share of at most 50\%. Destination addresses are not grouped. \item[P4, shared-class leave-one-source-out.] Train on three sources, test on the fourth, restricted to labels that satisfy the thresholds in the held-out source and in the other three together. \item[P4b, binary leave-one-source-out.] The same, with the label reduced to benign versus attack; all flows of the held-out source are tested. \item[P5, family leave-one-source-out.] The same on attack families (Section~\ref{ssec:taxonomy}) plus the benign class. A provisional taxonomy defined before the ATT\&CK grounding was kept as a sensitivity check. \end{description} In the leave-one-source-out protocols, 10\% of the training flows serve as validation set for early stopping. The validation set is never used for scoring. \begin{table}[t] \centering\small \caption{Protocols and evaluable labels (feature set A / B; for the leave-one-source-out protocols the counts are for UNSW-NB15 / ToN-IoT / BoT-IoT / CSE-CIC-IDS2018 as held-out source). Clean test flows are those without a leak flag, in labels evaluable under B.} \label{tab:protocols} \resizebox{\textwidth}{!}{\input{sections/tab_protocols}} \end{table} \subsection{What the protocols can evaluate} \Figref{fig:evaluability} shows the evaluability matrix summarised by protocol. \begin{figure}[t] \centering \includegraphics[width=0.78\textwidth]{fig_evaluability} \caption{Number of labels evaluable per protocol for feature sets A and B. P4 with CSE-CIC-IDS2018 held out has only the benign class and is not scored.} \label{fig:evaluability} \end{figure} Under P1, 31 labels are evaluable with feature set A and 26 with B; under P2, 30 and 26. For B, Brute\_Force\_-XSS, DoS\_attacks-SlowHTTPTest, FTP-BruteForce, SQL\_Injection, Theft and Worms drop out, because their distinct training vectors or clean test flows fall below the thresholds. Worms keep only 24 test flows after deduplication under P1 and Theft has no test window under P2. The rarest classes named in the dataset description (Worms with 158 flows, SQL\_Injection with 440 and Brute\_Force\_-XSS with 480) therefore cannot be scored under any valid protocol. The six remaining classes that we call the \emph{tail} are those with fewer than 5,000 training flows in the full P2 training split: Analysis, Brute\_Force\_-Web, DDOS\_attack-LOIC-UDP, Shellcode, ransomware and mitm, with 74 to 775 test flows each. P3 supports only three labels: the benign class, Brute\_Force\_-Web and Infilteration. The reason is that attacks originate from very few addresses: DoS has 19 distinct source IPs, DDoS 14 and Reconnaissance 12, and six classes (DoS\_attacks-Hulk, FTP-BruteForce, DoS\_attacks-SlowHTTPTest, DoS\_attacks-GoldenEye, SSH-Bruteforce, DoS\_attacks-Slowloris) have a single source IP. Holding out source addresses therefore cannot be evaluated for most attacks. The leave-one-source-out protocols cover four labels for UNSW-NB15, ToN-IoT and BoT-IoT as held-out source (benign, DoS, Reconnaissance and Backdoor; benign, DDoS, DoS and Backdoor; benign, DDoS, DoS and Reconnaissance, respectively), and only the benign class for CSE-CIC-IDS2018, which is why that fold is not scored. Family-level evaluation covers five, six, three and five labels for the four held-out sources. \subsection{Leakage inside the valid protocols} \Figref{fig:leakage} shows how much of each test set repeats the training set when identifiers are dropped. Under P1, 67\% of the test flows have a same-label duplicate of their B vector in the training set; under P2 the share is 34\% and under P3 20\%. For the binary leave-one-source-out protocols, 21\% of the ToN-IoT test flows and 4.8\% of the CSE-CIC-IDS2018 test flows have a same-label B duplicate in the other three sources, which means identical flows are shared between sources. Cross-label conflicts are rare in the three within-source protocols (below 1\% of test flows) and reach 4\% for ToN-IoT as held-out source. \begin{figure}[t] \centering \includegraphics[width=0.7\textwidth]{fig_leakage} \caption{Share of test flows that are clean, that duplicate a training flow with the same label on feature set B, or that conflict with training flows of another label.} \label{fig:leakage} \end{figure} \subsection{Attack families grounded in MITRE ATT\&CK} \label{ssec:taxonomy} Family-level evaluation needs a grouping of the 31 attack classes. We grounded it in MITRE ATT\&CK Enterprise version 20.0 \citep{mitre2026attack}. For each of the 36 pairs of source dataset and attack label, the behaviour described in the dataset papers \citep{sarhan2021netflow,koroniotis2019development,alsaedi2020ton,sharafaldin2017reliable} was extracted and mapped to one or more ATT\&CK techniques; identifiers, names and tactics were looked up in the ATT\&CK data file, and a second independent pass challenged each mapping. That pass flagged 27 of the 36 rows for correction, mostly the specificity of the technique and an over-stated confidence. The extraction and re-check were performed with language-model assistance and have not been validated against expert annotation. The mapping is by (source, label) because the same label can describe different behaviour in different sources. A label's family is the technique cluster of its primary technique; where sources disagree, the higher-confidence mapping is used, and Backdoor is resolved to command and control by an explicit rule because the UNSW-NB15 and ToN-IoT mappings differ and both have low confidence. \Tabref{tab:taxonomy} lists the resulting 11 families. Six mappings have high confidence, 14 medium and 16 low; the low ones are mostly UNSW-NB15 classes that the papers describe in a single line, and Generic (a cryptographic collision attack) has no ATT\&CK counterpart. ATT\&CK describes adversary behaviour, not flow statistics, so a family need not be separable in NetFlow features. The taxonomy differs from the provisional one in that Fuzzers join denial of service, MITM joins credential access, Analysis joins exploitation, and Worms, Shellcode, ransomware and Theft become singleton families (adjusted Rand index 0.69 over the 31 classes, 0.996 when weighted by flows). \begin{table}[t] \centering\small \caption{Family taxonomy v1 derived from ATT\&CK. H/M/L: number of source--label mappings with high, medium and low confidence per family.} \label{tab:taxonomy} \input{sections/tab_taxonomy} \end{table}