IG-OpenMax for UAV RF Open-Set Recognition

• The paper introduces a generative refinement layer that synthesizes simulated unknown features via a conditional DCGAN and retrains only the classification head to improve open-set detection.
• It integrates multi-domain supervised contrastive learning with advanced Weibull tail calibration to refine activation vectors for both known and unseen UAV RF signals.
• Empirical results on the DroneRFa dataset demonstrate a balanced performance, achieving 95.12% known and 96.08% unknown accuracy with minimal additional computation.

Searching arXiv for the primary paper and directly related methods to ground the article in current sources. {"query": "\"Multi-Domain Supervised Contrastive Learning for UAV Radio-Frequency Open-Set Recognition\" OR (Gao et al., 18 Aug 2025)", "max_results": 5} {"query":"(Gao et al., 18 Aug 2025)","max_results":10} Improved Generative OpenMax (IG-OpenMax) is an open-set recognition algorithm introduced within the multi-domain supervised contrastive learning framework for UAV radio-frequency (RF) surveillance in low altitude integrated sensing and communication (LA-ISAC) networks. In the reported formulation, IG-OpenMax extends standard OpenMax by adding a feature-level generative component based on a conditional DCGAN, filtering generated samples through a pretrained closed-set network, and retraining only the classification head while freezing the feature extractor. It is used to construct an open-set recognition model, Open-RFNet, for distinguishing known UAV types from invasive or previously unseen UAVs, and the reported system achieves 95.12% in closed-set and 96.08% in open-set under 25 UAV types (Gao et al., 18 Aug 2025).

1. Operational setting and placement within UAV RF open-set recognition

The algorithm is presented for UAV RF open-set recognition in the context of 5G-Advanced-enabled LA-ISAC networks, where non-cooperative UAVs constitute a security threat because unauthorized flight incidents occur frequently. The broader framework first fuses texture features and time-frequency position features from a ResNet and a TransformerEncoder, and then applies supervised contrastive learning to optimize the feature representation of the closed-set samples. IG-OpenMax is the component introduced to surveil invasive UAVs that appear in real life, after the closed-set representation has been learned (Gao et al., 18 Aug 2025).

Within that pipeline, IG-OpenMax serves as the mechanism that converts a closed-set classifier into an open-set model. The paper describes this resulting model as Open-RFNet. Its design premise is that closed-set training alone is insufficient when unknown UAV types must be rejected or assigned to an unknown class. A plausible implication is that IG-OpenMax should be read not as a standalone backbone, but as an open-set refinement layer built on top of a feature extractor already optimized by multi-domain supervised contrastive learning.

2. OpenMax foundation: activation-vector calibration and the unknown class

IG-OpenMax is defined against the baseline of standard OpenMax, which augments a normal softmax-based classifier with an “unknown” class by calibrating final logits, or activation vectors (AVs), using Weibull cumulative distributions fitted on the tails of each known class (Bendale et al., 2015).

Let a trained network produce for an input $x$ an activation vector

$v(x) = (v_0(x), v_1(x), \dots, v_{N-1}(x)),$

where $N$ is the number of known classes. For each class $j \in \{0,\dots,N-1\}$, the correctly classified training AVs are collected as

$S_j = \{v_i(x_i)\mid \hat{y}_i = y_i = j\},$

and the mean activation vector (MAV) is computed as

$\mu_j = \operatorname{mean}(S_j).$

The distribution of distances

$D_j = \{\|v-\mu_j\|_2 : v\in S_j\}$

is then modeled by fitting a three-parameter Weibull distribution

$$F_w(d;\lambda_j,k_j,\chi_j)=1-\exp\left[-\left(\frac{d-\chi_j}{\lambda_j}\right)^{k_j}\right],$$

where $\lambda_j>0$ is the scale, $k_j>0$ the shape, and $v(x) = (v_0(x), v_1(x), \dots, v_{N-1}(x)),$0 the location; in practice, only the largest $v(x) = (v_0(x), v_1(x), \dots, v_{N-1}(x)),$1 distances, the tail, are fitted via EVT (Gao et al., 18 Aug 2025).

At inference time, if $v(x) = (v_0(x), v_1(x), \dots, v_{N-1}(x)),$2 are the indices of classes sorted by $v(x) = (v_0(x), v_1(x), \dots, v_{N-1}(x)),$3’s magnitude in descending order, then for the top $v(x) = (v_0(x), v_1(x), \dots, v_{N-1}(x)),$4 classes a multiplicative survival weight is computed:

$v(x) = (v_0(x), v_1(x), \dots, v_{N-1}(x)),$5

The recalibrated AV is

$v(x) = (v_0(x), v_1(x), \dots, v_{N-1}(x)),$6

and the unknown logit is derived from the mass subtracted from known classes:

$v(x) = (v_0(x), v_1(x), \dots, v_{N-1}(x)),$7

A SoftMax over the $v(x) = (v_0(x), v_1(x), \dots, v_{N-1}(x)),$8 entries $v(x) = (v_0(x), v_1(x), \dots, v_{N-1}(x)),$9 yields the final open-set probabilities. In IG-OpenMax, this calibration logic remains the final decision stage, but it is applied after a different procedure for constructing the AV space.

3. Limits of vanilla OpenMax in the UAV RF regime

The stated motivation for IG-OpenMax is that vanilla OpenMax is not sufficiently robust for UAV RF signals. The paper identifies two specific limitations. First, UAV RF signals are highly non-stationary and may employ frequency hopping; as a result, the high-level AV space of a closed-set CNN can still be quite tight, and fitting Weibull tails on known classes alone often fails to push truly unseen UAV types into the unknown slot. Second, noisy channels and limited training examples of new UAVs limit OpenMax’s ability to distinguish malicious or invasive UAVs without explicit examples (Gao et al., 18 Aug 2025).

These constraints are specific to the signal regime rather than merely to classifier architecture. The issue is not only whether the final classifier can be recalibrated, but whether the representation contains enough structure for unknowns to occupy separable regions. This suggests that the paper’s central modification is aimed at reshaping the effective decision boundary in AV-space rather than replacing Weibull calibration itself.

4. IG-OpenMax algorithm: generative simulation, filtering, and head-only retraining

IG-OpenMax enhances OpenMax by explicitly synthesizing “unknown” feature vectors via a conditional DCGAN (cDCGAN), filtering them through the pretrained closed-set network, and then retraining only the classification head while freezing the feature extractor. The reported effect is to yield simulated unknowns that better approximate real unseen features and improve decision boundaries in AV-space (Gao et al., 18 Aug 2025).

The generative module is a cDCGAN with a generator $N$0 and critic $N$1. The generator is a deep convolutional network taking as input a noise vector $N$2 and a one-hot encoding of a known class label $N$3, and outputs a synthetic RF-signal sample $N$4. The critic is a deep convolutional network mapping $N$5. Training uses the WGAN-GP objective (Gulrajani et al., 2017):

$N$6

where $N$7 is the real-RF data distribution, $N$8 the generator distribution, $N$9 enforces the Lipschitz penalty, and $j \in \{0,\dots,N-1\}$0 is a weighting term (Gao et al., 18 Aug 2025).

The simulation of unknown features proceeds in four steps. First, the cDCGAN $j \in \{0,\dots,N-1\}$1 is trained on known-class data. Second, many samples $j \in \{0,\dots,N-1\}$2 are drawn under each known label $j \in \{0,\dots,N-1\}$3. Third, each $j \in \{0,\dots,N-1\}$4 is passed through the frozen closed-set network consisting of the feature extractor $j \in \{0,\dots,N-1\}$5 and classifier $j \in \{0,\dots,N-1\}$6. Fourth, the samples whose predicted label satisfies $j \in \{0,\dots,N-1\}$7 are collected as simulated unknowns $j \in \{0,\dots,N-1\}$8 (Gao et al., 18 Aug 2025).

The retraining strategy is explicitly two-stage. In Stage I, the full network $j \in \{0,\dots,N-1\}$9 is trained end-to-end on known classes only; the paper notes that this can be done, for example, with cross-entropy or supervised contrastive plus cross-entropy. In Stage II, the feature extractor $S_j = \{v_i(x_i)\mid \hat{y}_i = y_i = j\},$0 is frozen, the output dimension of $S_j = \{v_i(x_i)\mid \hat{y}_i = y_i = j\},$1 is expanded to $S_j = \{v_i(x_i)\mid \hat{y}_i = y_i = j\},$2, and $S_j = \{v_i(x_i)\mid \hat{y}_i = y_i = j\},$3 is trained on the union of known examples $S_j = \{v_i(x_i)\mid \hat{y}_i = y_i = j\},$4 with standard cross-entropy:

$S_j = \{v_i(x_i)\mid \hat{y}_i = y_i = j\},$5

After this head-only retraining, OpenMax recalibration is applied again on the AVs $S_j = \{v_i(x_i)\mid \hat{y}_i = y_i = j\},$6 to obtain the final probabilities over $S_j = \{v_i(x_i)\mid \hat{y}_i = y_i = j\},$7 known plus one unknown class (Gao et al., 18 Aug 2025).

The pseudo-code given for IG-OpenMax is correspondingly direct: input the known-class training set $S_j = \{v_i(x_i)\mid \hat{y}_i = y_i = j\},$8 and Weibull tail size $S_j = \{v_i(x_i)\mid \hat{y}_i = y_i = j\},$9; train $\mu_j = \operatorname{mean}(S_j).$0 on $\mu_j = \operatorname{mean}(S_j).$1 to obtain “Open-RFNet-B”; train cDCGAN $\mu_j = \operatorname{mean}(S_j).$2 on $\mu_j = \operatorname{mean}(S_j).$3 with WGAN-GP; generate many samples $\mu_j = \operatorname{mean}(S_j).$4 and collect misclassified ones as $\mu_j = \operatorname{mean}(S_j).$5; freeze $\mu_j = \operatorname{mean}(S_j).$6, expand $\mu_j = \operatorname{mean}(S_j).$7 to $\mu_j = \operatorname{mean}(S_j).$8 outputs, and fine-tune $\mu_j = \operatorname{mean}(S_j).$9 on $D_j = \{\|v-\mu_j\|_2 : v\in S_j\}$0; fit Weibull tails on the new AVs of known classes; and at test time compute $D_j = \{\|v-\mu_j\|_2 : v\in S_j\}$1, apply OpenMax weights $D_j = \{\|v-\mu_j\|_2 : v\in S_j\}$2, form $D_j = \{\|v-\mu_j\|_2 : v\in S_j\}$3, derive $D_j = \{\|v-\mu_j\|_2 : v\in S_j\}$4, and apply SoftMax (Gao et al., 18 Aug 2025).

5. Calibration stage, computational profile, and architectural cost

After head retraining, IG-OpenMax returns to the standard OpenMax calibration stage. The paper states that, after retraining the head $D_j = \{\|v-\mu_j\|_2 : v\in S_j\}$5, OpenMax recalibration is applied once again on the AVs to obtain the final probabilities over $D_j = \{\|v-\mu_j\|_2 : v\in S_j\}$6 known plus one unknown class. In that sense, the improvement does not discard Weibull-based tail modeling; it changes the AV geometry on which that modeling operates (Gao et al., 18 Aug 2025).

The computational complexity is described as modest relative to backbone inference. The backbone of IG-OpenMax is the same feature extractor $D_j = \{\|v-\mu_j\|_2 : v\in S_j\}$7 used in the closed-set system, specifically a ResNet-18 plus TransformerEncoder. Adding the generative refinement only involves retraining the small classification head $D_j = \{\|v-\mu_j\|_2 : v\in S_j\}$8 and fitting Weibull distributions, which is $D_j = \{\|v-\mu_j\|_2 : v\in S_j\}$9 for MAV computation and $$F_w(d;\lambda_j,k_j,\chi_j)=1-\exp\left[-\left(\frac{d-\chi_j}{\lambda_j}\right)^{k_j}\right],$$0 per test sample for recalibration. Compared to vanilla OpenMax, the extra cost is dominated by generating $$F_w(d;\lambda_j,k_j,\chi_j)=1-\exp\left[-\left(\frac{d-\chi_j}{\lambda_j}\right)^{k_j}\right],$$1 samples from $$F_w(d;\lambda_j,k_j,\chi_j)=1-\exp\left[-\left(\frac{d-\chi_j}{\lambda_j}\right)^{k_j}\right],$$2 and filtering them through $$F_w(d;\lambda_j,k_j,\chi_j)=1-\exp\left[-\left(\frac{d-\chi_j}{\lambda_j}\right)^{k_j}\right],$$3, one forward pass per sample, and by retraining the head $$F_w(d;\lambda_j,k_j,\chi_j)=1-\exp\left[-\left(\frac{d-\chi_j}{\lambda_j}\right)^{k_j}\right],$$4, which is described as negligible if $$F_w(d;\lambda_j,k_j,\chi_j)=1-\exp\left[-\left(\frac{d-\chi_j}{\lambda_j}\right)^{k_j}\right],$$5 is small (Gao et al., 18 Aug 2025).

The implementation figures reported in the paper reinforce that distribution of cost. The combined $$F_w(d;\lambda_j,k_j,\chi_j)=1-\exp\left[-\left(\frac{d-\chi_j}{\lambda_j}\right)^{k_j}\right],$$6 system, using ResNet plus Transformer, has approximately 205 M parameters and about 1.03 GFLOPs per forward pass, whereas ResNet alone has 0.7 M parameters and 0.47 GFLOPs. The paper therefore states that the bulk of the cost remains in feature extraction, not in IG-OpenMax’s extra stages (Gao et al., 18 Aug 2025).

6. Empirical behavior on DroneRFa and comparative performance

The experiments are conducted with a large-scale UAV open dataset, identified in the detailed description as the DroneRFa dataset. In the closed-set setting with 20 known UAV types, the pre-refinement network achieves approximately 99.4% top-1 accuracy on known validation samples. In the open-set setting with 20 known and 5 held-out unknown classes, the post-IG-OpenMax system reports a Known Accuracy Rate (KAR) of 95.12% and an Unknown Accuracy Rate (UAR) of 96.08%, giving a performance gap $$F_w(d;\lambda_j,k_j,\chi_j)=1-\exp\left[-\left(\frac{d-\chi_j}{\lambda_j}\right)^{k_j}\right],$$7 (Gao et al., 18 Aug 2025).

Method	KAR	UAR
Open-RFNet-B	95.23%	93.84%
Open-RFNet-G	95.29%	93.12%
IG-OpenMax	95.12%	96.08%
S3R	90.86%	99.2%
UIOS	95.12%	93.92%

The comparative interpretation reported in the paper is specific. Relative to vanilla OpenMax, referred to as “Open-RFNet-B,” IG-OpenMax raises UAR by $$F_w(d;\lambda_j,k_j,\chi_j)=1-\exp\left[-\left(\frac{d-\chi_j}{\lambda_j}\right)^{k_j}\right],$$8 while changing KAR from 95.23% to 95.12%. Relative to Generative OpenMax, referred to as “Open-RFNet-G,” IG-OpenMax raises UAR by $$F_w(d;\lambda_j,k_j,\chi_j)=1-\exp\left[-\left(\frac{d-\chi_j}{\lambda_j}\right)^{k_j}\right],$$9 while changing KAR from 95.29% to 95.12%. Against two recent RF open-set methods, S3R and UIOS, IG-OpenMax yields the smallest reported KAR–UAR gap, 0.96%, compared with 8.34% for S3R and 1.20% for UIOS (Gao et al., 18 Aug 2025).

These results are presented as evidence that the method improves the balance between recognizing known classes and rejecting unknown ones. A plausible implication is that the paper values not only high UAR or high KAR in isolation, but also gap minimization as a stability criterion for open-set deployment.

7. Interpretation, relation to prior variants, and recurring misconceptions

The paper provides four reasons for why IG-OpenMax works. First, by generating and filtering at the feature level, simulated unknowns lie close to the true unknown AVs and push decision boundaries outward. Second, freezing $\lambda_j>0$0 avoids disturbing the multi-domain features learned by the contrastive-plus-Transformer-augmented backbone. Third, retraining only the light head $\lambda_j>0$1 is efficient and prevents overfitting on limited unknown samples. Fourth, Weibull calibration on this refined feature space yields tighter separation between known clusters and truly open samples (Gao et al., 18 Aug 2025).

In relation to earlier variants, the summary states that IG-OpenMax combines the benefits of explicit unknown sample synthesis, as in G-OpenMax, with a feature-level, head-only retraining strategy. The reported outcome is more realistic unknown AVs, sharper class boundaries, and higher open-set detection rates with minimal overhead (Gao et al., 18 Aug 2025). This places IG-OpenMax within the family of OpenMax-derived methods while distinguishing it by where the generative mechanism enters the pipeline: not merely as data augmentation, but as a controlled source of simulated unknowns selected through misclassification by the pretrained network.

A common misconception would be to treat IG-OpenMax as a replacement for OpenMax calibration. The reported method does not remove the Weibull-tail procedure; instead, it performs final open-set calibration “as in Sec. 1.3” after retraining the classification head. A second misconception would be to interpret it as a full-network open-set finetuning scheme. The algorithm explicitly freezes the feature extraction layers and retrains only the classification layer according to the unknown samples. In the paper’s formulation, these two constraints are central rather than incidental: the feature extractor remains fixed, and the recalibration stage remains OpenMax-based (Gao et al., 18 Aug 2025).