The proliferation of vulnerable Internet-of- Things (IoT) devices has enabled large-scale cyberattacks. Solutions like Hestia and HomeSnitch have failed to comprehensively address IoT security needs. This research evaluates if Wireguard, an emerging VPN protocol, can provide efficient security tailored for resource-constrained IoT systems. We compared Wireguard's performance against standard protocols Open VPN and IPsec in a simulated IoT environment. Metrics measured included throughput, latency, and jitter during file transfers. Initial results reveal Wireguard's potential as a lightweight yet robust IoT security solution despite disadvantages for Wireguard in our experimental environment. With further testing, Wireguard's simplicity and low overhead could enable widespread VPN adoption to harden IoT devices against attacks. The protocol's advantages in setup time, performance, and compatibility make it promising for integration especially on weak IoT processors and networks.
The published MLP-based DRL in finance has difficulties in learning the dynamics of the environment when the action scale increases. If the buying and selling increase to one thousand shares, the MLP agent will not be able to effectively adapt to the environment. To address this, we designed a CNN agent that concatenates the data from the last ninety days of the daily feature vector to create the CNN input matrix. Our extensive experiments demonstrate that the MLP-based agent experiences a loss corresponding to the initial environment setup, while our designed CNN remains stable, effectively learns the environment, and leads to an increase in rewards.
The shift to multimodal Vision-Language Models (VLMs) theoretically promises improved robustness against visual data leakage vulnerabilities of convolutional architectures via aligning retinal pathology with existing medical semantics. Nevertheless, zero-shot application assumes that these models achieve genuine cross-modal reasoning rather than shallow text token-matching. To differentiate visual-semantic reasoning from gradient-driven memorization, we conducted strictly-frozen inference on the FLAIR retinal foundation model using 1,744 MESSIDOR-2 fundus images. We developed an automated high-concurrency Unified Medical Language System (UMLS) sanitization pipeline for enforcing absolute semantic neutrality and ensuring the removal of latent medical ontologies from both lay language prompts and exclusionary evaluation texts. Semantic perturbation resulted in cross-modal collapse. While baseline prompts limited only to labels yielded 56.18% accuracy (Expected Calibration Error [ECE]: 4.45%), UMLS-sanitization of expert terms to lay language dropped the performance to 29.62% (ECE: 38.70%). Further application of negative logical operators dropped the accuracy to 26.90% (ECE: 36.15%), resulting in the emergence of a 24.0% class-conditional false positive rate for Proliferative Diabetic Retinopathy diagnosis among completely normal eyes merely due to the presence of pathological text tokens. Current medical Vision-Language Models (VLMs) suffer significantly from lexical overfitting and operate as fragile bags-of-words rather than reasoning agents. Under semantic perturbation, they exhibit epistemic arrogance -- a potentially fatal property in a clinical context that allows the model architecture to hide biological illiteracy from the end-user under a facade of mathematically unjustified statistical overconfidence.