Where Post-Quantum Migration Hurts a Wi-Fi IoT Endpoint: ESP32 Costs, TLS 1.3 Wire Overhead, and Deployment Guidelines

This repository is the artifact for the paper:

K. Malinka, M. Perešíni, R. Hranický, A. Firc, A. Smrčka, F. Bučko, O. Hujňák, J. Kratochvíl, and G. Sitko, "Where Post-Quantum Migration Hurts a Wi-Fi IoT Endpoint: ESP32 Costs, TLS 1.3 Wire Overhead, and Deployment Guidelines," in Proc. IEEE NCA 2026 Workshops (ICON 2026), 2026.

The paper measures two things. The first is the CPU, RAM, and flash cost of post-quantum primitives on an ESP32 (Xtensa LX6). The second is the size of TLS 1.3 handshakes on a Raspberry Pi and Linux testbed.

Contents

Path	Content
data/esp32_primitives.csv	Cycles (min, mean, max over 20 runs), peak RAM (static, stack, heap), and incremental flash (DRAM, IRAM, code) for each operation.
data/tls13_message_sizes.csv	TLS 1.3 record bytes per handshake message for 3 configurations and 2 authentication modes.
data/object_sizes.csv	Key, ciphertext, and signature sizes of the measured parameter sets and their standardized successors.
analysis/derived_models.py	Compute the numbers from data for figures and analysis.
firmware/	ESP32 measurement firmware (ESP-IDF v5.2, FreeRTOS). See the note below.
tls/	wolfSSL TLS 1.3 client and server, certificate scripts, and captures. See the note below.

Run analysis

Need Python 3.8 or later. The script uses the standard library.

python analysis/derived_models.py

Measurement environment

Primitive measurements:

• Board: ESP32-DevKitC V4 with ESP32-WROOM-32E (dual-core Xtensa LX6, 240 MHz, 520 KiB SRAM, 4 MiB flash).
• Firmware: ESP-IDF v5.2, FreeRTOS, compiler option -Os, bootloader log level "warning", assertions silent.
• Post-quantum code: PQClean clean implementations. The random-bytes adapter calls esp_fill_random.
• Classical code: wolfSSL (WOLFSSL_ESP32, HAVE_CURVE25519, HAVE_ED25519, USE_FAST_MATH, WOLFSSL_SMALL_STACK).
• Method: each operation runs 20 times in a new high-priority task on core 1. Background tasks run on core 0. An empty-task baseline is subtracted. Cycles come from xthal_get_ccount, stack from uxTaskGetStackHighWaterMark, heap from the ESP-IDF heap hooks, and flash from idf.py size and esp_idf_size.

TLS measurements:

• Client: Raspberry Pi 4 Model B (Raspberry Pi OS). Server: Dell Latitude 5521 (Ubuntu 22.04 LTS). Switch: MikroTik hAP ax3, Ethernet.
• TLS stack: wolfSSL post-quantum examples at commit 20d13d8526f5435f2a6a5c637702bf9170d7268e, liboqs 0.8.0, and the OQS OpenSSL 1.1.1 fork (OQS-OpenSSL_1_1_1-stable) for certificates.
• wolfSSL configuration: ./configure --enable-tls13 --enable-ed25519 --enable-curve25519 --disable-rsa --enable-sha3 --enable-sha512 --enable-experimental --with-liboqs CFLAGS="-DHAVE_SECRET_CALLBACK"
• PKI: root CA (preloaded, not sent), intermediate CA, and end-entity certificates, all in the signature family of the configuration.
• Cipher suite: TLS_AES_256_GCM_SHA384. The client offers 12 groups and 9 signature algorithms.
• Captures: OQS Wireshark build with wolfSSL OIDs, capture filter port 11111 and host <ip>, decryption with the client key log.
• The byte counts include TLS record headers. They do not include Ethernet, IP, or TCP headers.

Data

The data/ come from the measurement results.