Tracking the Slope, Not the Level: Peer-Adjusted Behavioral Trajectories Against Slow Insider Evasion

Insider threat detectors judge each user-day by how far it sits from a reference. That reference is either the user's own recent history or the user's role peers. Patient insiders can beat both references. An employee who raises their exfiltration a little each week slowly moves their own history along with them, which practitioners call the "boiling frog" tactic. A department-wide reorganization or security incident does the opposite and makes every peer look anomalous at the same moment. Clinical medicine has faced the same two problems with blood markers for decades. A single lab value says little. What predicts outcomes is the direction and rate of change of a patient's marker relative to what is expected for patients like them, plus the cumulative exposure that keeps risk elevated even after the marker comes back down.

We propose to model each user's behavior the way a longitudinal biomarker study models a patient. Each daily activity count is split into three parts: a trend shared by the user's role peers, a stable personal offset, and a slowly varying personal drift. Only the drift is scored, using its slope and its accumulated burden. The claim is concrete. Taking out the shared trend before tracking the personal slope should catch slow-ramp insiders that current detectors miss at a fixed alert budget. It should also keep false alarms flat during organization-wide shifts, which inflate false alarms for peer-based detectors.

Core Claim

The open problems listed for insider detection include concept drift from role changes, the gap between synthetic attacks and adaptive human adversaries, and sensitivity of peer-based neighborhoods. All of them reduce to one confusion: the detectors do not separate whose change a deviation is. The diachronic-plus-parallel scheme of A Data-Driven Evaluation for Insider Threats compares a user's activity vector over discrete audit periods against their history with a bounded divergence, and against peers with local outlier density. It scores levels in each period and treats risk as growing with distance. As a result, a gradual personal ramp disappears into the moving history, and a shared shock shows up in the peer comparison as everyone being far from everyone. We argue that the relevant quantity is the user's idiosyncratic rate of change after peer adjustment, integrated over time. We expect this quantity to separate slow malicious escalation from both benign power-user volume and benign shared drift.

What Carries Over From Biomarker Tracking

The biomarker papers in our file do three distinct jobs for this proposal: they motivate it, they constrain it, and they supply a method.

They motivate it because several cohort studies find that trajectories beat snapshots. Association Between High-Sensitivity C-Reactive Protein Trajectories and the Incidence of Metabolic Syndrome clusters 4,312 patients' inflammatory-marker paths into low-stable, increasing and decreasing classes. Even the decreasing class carried elevated hazard (adjusted HR 1.408 against low-stable), which the authors call a legacy effect. Cross-cohort insights into the association of handgrip strength transitions and burdens with cardiovascular disease risk likewise finds that transitions and cumulative burden predict risk beyond the baseline value. Both results motivate scoring cumulative positive drift rather than the current day, so that an insider who exfiltrates and then goes quiet is not immediately cleared. That gap is real in current systems: daily-likelihood scoring as in Unsupervised User-Based Insider Threat Detection Using Bayesian Gaussian Mixture Models has no such memory.

They constrain it through a negative result. The association between the neutrophil-to-lymphocyte ratio and mortality in critical illness shows that a ratio that predicts mortality across ICU patients loses all association in sepsis, where systemic disruption saturates it. We take this as a warning about any score defined relative to a cohort. When the whole cohort shifts, a peer-relative level carries no information. The fix is to model the shared component explicitly rather than hope peers stay stable. We also take a constraint from the tension the research file identifies. Simon et al. argue that one population model does not fit an accountant and an administrator alike. Rigid population trajectory classes of the kind used for hsCRP would bring back the "active benign" problem. We therefore borrow partial pooling from mixed-effects longitudinal modeling: each user has their own parameters, shrunk toward their role. We do not borrow fixed population classes.

They supply the method. Clinical trajectory work must cope with irregular, missing measurements; the albumin study Association between albumin changes and prognosis in older patients with acute myocardial infarction excluded nearly 750 patients for missing values. A state-space trend model handles missing days natively. That matters for employees on leave, part-timers and accounts with sparse activity. The albumin study's restricted cubic splines, which are smooth piecewise-cubic curves with linear tails, found a non-monotone change–risk relation. We use the same tool only as a diagnostic, mapping drift scores to observed attack rates to check whether risk rises monotonically. We do not use splines as the detector, because their knot sensitivity is a known weakness.

On the security side, the closest work treats logs as signals but scores levels or learned representations. Log2Sig turns logs into multichannel frequency signals, decomposes them into slow and fast modes, and classifies with a Mamba encoder; it never peer-adjusts or tracks slope. MambaITD smooths inter-event intervals with an exponential moving average and sets per-user thresholds with Otsu's method, again without a velocity term. Graph Based Framework for Malicious Insider Threat Detection adds Big Five personality scores as static attributes and itself points toward temporal modeling as future work. None of these subtracts a role-shared trend or scores accumulated idiosyncratic drift.

Research Program

For each user $u$, behavioral channel $c$ (after-hours logons, removable-device events, file copies, external email, job-search and cloud-storage web categories) and day $t$, we model the log-transformed count as

$$y_{u,c,t} = m_{r(u),c}(t) + \alpha_{u,c} + \delta_{u,c}(t) + \varepsilon_{u,c,t},$$

Here $m_{r(u),c}(t)$ is the day-$t$ trend shared by the user's role $r(u)$, estimated robustly across role members. $\alpha_{u,c}$ is the user's stable offset, shrunk toward the role mean. $\delta_{u,c}(t)$ is a personal local linear trend, meaning a level plus a slope that each move as a random walk and are estimated by a Kalman filter that simply skips missing days. $\varepsilon$ is day-level noise. The alert score combines two quantities: the posterior probability that the user's current slope exceeds the spread of slopes among role peers, and a burden term equal to the discounted sum of positive $\delta$ over a trailing window. The discount rate controls how long a past excursion keeps risk elevated, which is the legacy effect.

The program has three parts. The first is building this model and its partial-pooling estimator for roles of very different sizes. The second is constructing evaluation conditions that existing benchmarks lack, namely slow-ramp attacks and shared shocks, described below. The third is ablating the model so that we learn which ingredient matters: peer adjustment, slope, burden, or partial pooling.

One limitation we will measure rather than assume away concerns individual role changes such as promotions. They are not shared with peers and will look like idiosyncratic drift. We will report what fraction of false alarms coincide with such changes wherever the benchmark's organizational records allow us to identify them.

Baselines

The naive reference point is a static per-user z-score against a frozen first-month baseline, thresholded per channel. It is the rule-based snapshot a security team would deploy by default. Any method that cannot beat it under slow ramps has no claim.

The strongest existing methods differ by benchmark. On CERT r4.2, Log2Sig reports F1 0.956 and precision 0.990. Unsupervised User-Based Insider Threat Detection Using Bayesian Gaussian Mixture Models reports AUC 0.958 at 6.9% false positive rate. RedChronos reports detection rate 0.987 at false positive rates of 0.019–0.022. Detect Insider Threat with Associated Session Graph reports 99% F1. These numbers come from different splits and units (user-day, session, user), so we will not compare against them as printed. We will rerun the BGMM detector and Log2Sig ourselves, on our splits and with the same hyperparameter search budget, because both are unsupervised or lightly supervised and publicly described well enough to reimplement. These two are the strongest competitors because one is the best-documented personal-baseline method and the other already models slow components of behavioral signals. We will also run the dual-baseline LOF of A Data-Driven Evaluation for Insider Threats as the representative peer-comparison method, since it is the method the shared-shock condition targets. On the human red-team data, Detection of Masqueraders Based on Graph Partitioning of File System Access Events reports AUC 0.851 on TWOS. For slow-ramp variants of any benchmark there is no reported result on Lacuna, so all numbers there will be ours.

The key ablation is role-demeaned CUSUM. CUSUM is a cumulative-sum change detector that accumulates evidence of a sustained shift. Here we run it per user on counts after subtracting the daily role median. This strips the proposal down to its central ingredient, peer adjustment followed by accumulation, without the state-space model or partial pooling. Further ablations remove the role trend, the slope term and the burden term one at a time. Every method gets the same features, the same training window, the same tuning budget (a fixed number of configurations selected on a held-out validation month) and the same alert budget.

Beating a baseline means one of two things. Either we achieve strictly higher detection of ramped insiders at a matched alert budget of 1% of user-days, with confidence intervals from user-level bootstrap. Or, under shared shocks, our false-alarm rate stays within its no-shock value while the baseline's rises. On unmodified CERT, the bar is non-inferiority, meaning no loss beyond bootstrap noise.

Why not just…?

Why not just run CUSUM or Page–Hinkley change detection on each user's own history? A sustained shift against a frozen reference will be caught. Unadjusted CUSUM, however, fires equally on a department-wide workload rise, so it should fail exactly where the sepsis analogy predicts. The shared-shock experiment measures this directly. The peer-adjusted version of CUSUM is our main ablation, not a strawman, and we treat it as a genuine threat. If it matches the full model everywhere, we will report that the contribution is the peer-adjusted slope-and-burden formulation and the evaluation protocol, and drop the state-space machinery. We expect the two to differ in small roles and for users with many missing days, where a raw role median is noisy and a CUSUM statistic has no principled way to bridge gaps. We will report results stratified by role size and activity sparsity to test this.

Why not just add differenced features or rolling slopes to an existing detector such as Isolation Forest or the BGMM? Adding slopes is cheap, and we will run it as a baseline. It still scores slopes against the user's own or the global distribution, without removing the role-shared trend. Shared shocks therefore still produce correlated false alarms, and the model has no memory once slopes return to zero. The ramp and shock experiments test whether slopes alone recover most of the gain. If they do, the burden and pooling terms are not earning their place.

Why not just retrain or adapt the detector more often to handle drift? Faster adaptation is precisely what a boiling-frog insider exploits. The faster the baseline follows the user, the more a slow ramp is absorbed. We will sweep the retraining interval of the BGMM baseline and plot ramp detection against benign-drift false alarms. We expect a trade-off curve that no interval escapes, whereas the decomposition should sit above it.

Proposed Experiments

The first experiment is a no-regression check on CERT r4.2 and r5.2 with their standard injected scenarios, using a temporal split and user-day scoring. These attacks are mostly abrupt. Here the proposal must only stay within bootstrap noise of the BGMM and Log2Sig reruns.

The second experiment is the decisive slow-ramp condition. We take each CERT malicious user's injected actions and redistribute them over 4, 8 and 12 weeks with a linearly increasing daily intensity, keeping total volume fixed. The question is whether detection at a fixed 1% alert budget, and the delay until first alert, degrade more slowly with ramp length for the proposal than for the BGMM, Log2Sig and LOF reruns. If peer-adjusted slope tracking is the right idea, baselines should lose recall as ramps lengthen while the proposal's recall falls much less. If the curves are parallel, the core claim is wrong.

The third experiment is a shared-shock condition. We multiply selected channels for every member of one department by a factor for two weeks, mimicking a reorganization or incident response, both with and without concurrent ramped insiders in that department. The test is whether the proposal's false-alarm rate stays near baseline while the peer-LOF and unadjusted CUSUM false-alarm rates rise. We will also check whether ramped insiders inside the shocked department remain detectable.

The fourth experiment repeats the ramp and shock injections on SALE, a real enterprise log with 6,886 users across 30 roles. There, A Data-Driven Evaluation for Insider Threats reports 100% accuracy with optimized LOF on abrupt injected anomalies. SALE is where role size varies most, so it is where partial pooling should beat role-demeaned CUSUM if it beats it anywhere.

The fifth experiment is TWOS, with human masqueraders, as a check that the method does not degrade on real adversaries against the reported AUC of 0.851. Masquerades there are short sessions rather than ramps, so we expect parity rather than gains and will say so.

Finally, we will fit restricted cubic splines of attack rate against the burden score on held-out data. This tests whether risk is monotone in burden or shows a plateau, which would tell us whether a single threshold is appropriate or whether the score needs recalibration.