个人阅读论文对专业词汇摘录
A
Attention
- Attention mechanisms:注意机
- Self-attention
Artificial Neural Networks:人工神经网络
Adversarial Network:对抗网络
adversarial examples:对抗例子
adversarial samples:对抗样本
annotation:标注,注释
authentication:身份验证
B
blockchain:区块链
backward propagation:反向传播
- backpropagation
- forward propagation:前向传播
blackbox:黑盒
将系统看成黑盒,内部如何实现是不需要了解的,只需要知道输入和预期输出。
binary:二进制
back-propagation:反向传播
Bluetooth:蓝牙
C
convolution:卷积
- Convolutional Neural Network (CNN)
classification:分类
contrastive learning:对比学习
cipher:密码
- ciphertext:密码本
code
- Encoders:编码器
- Decoders:解码器
correct:正确的,无误的
constant:常数,不变的
D
detection:检测
- object detection:目标检测
distillation:蒸馏
discriminative:判别
- discriminative task:判别任务
deblurring:去模糊
domain:领域
Data Augmentation:数据增强
downsampled:降采样
decision tree:决策树
Differential privacy (DP) , the gold standard for privacy protection, is considered an effective approach to defend against inference attacks. It requires that the privacy algorithms ensure that changes in individual records have little impact on the analysis results. Since DP guarantees that analyzers have difficulty distinguishing whether a given record is in datasets.
discretization:离散化,常用于数据离散化,老技术了,例如一个点离散到一个区域。
E
Ensemble Learning:集成学习
explanation techniques:解释技术
F
feasibility:可行性
fidelity:忠诚,忠实; 逼真; 保真度; 尽责,准确度
Federated Learning:联邦学习
fully-connected layers(FC):全连接层
Feed Forward Neural Network(FFN):前馈神经网络
G
Generative Adversarial Networks (GANs):生成对抗网络
gradient descent:梯度下降
feed-forward:前馈
H
hyper-parameter:超参数
hierarchical learning:层次学习
** hacker**:黑客
I
interpretability:可解释性
identification:身份认证
Inference attack, which aims to infer private information from data analysis tasks, is a significant threat to privacy-preserving data analysis. Many kinds of inference attacks have been proposed, such as Membership Inference Attacks (MIAs) , attribute inference attacks, and category inference attacks.
J
K
L
Linear regression:线性回归
M
multi-:大量、许多
module:模块,组件
malware:n. 恶意软件;恶意软件。malware是malicious software的合成形式;流氓软件
matrix:矩阵
memory:内存,存储器
METHODOLOGY(methodology):方法论
Membership inference attacks are familiar attacks on training samples in machine learning. It aims to infer whether a record is a training sample or not.
Multimodal:多模态
N
neuron:神经元
natural language processing (NLP)
O
P
pooling:池化
pixel:像素
parallel:并行
proof-of-concept:概念验证
pricacy protection: 隐私保护
probability density function (PDF):概率密度函数
Q
R
Recurrent:循环,递归
- Deep Recurrent Neural Networks:深度递归神经网络
reinforcement learning:强化学习
recognition:识别
- object recognition:对象识别
- facial recognition
- biometric recognition
Regularized:正则化
- Regularized iterative algorithms:正则化迭代算法
robustness:健壮性
resolution:分辨率
Randomized Response:随机响应,差分隐私中常用的一种方法
Reinforcement learning from human feedback (RLHF): 根据人类反馈的强化学习
S
state-of-the-art:使用最先进技术的;体现最高水平的
sub-field:子领域
semi:半
supervise:监督
- supervised learning
- semi-supervised :半监督
- unsupervised
semantic textual similarity (STS):语义文本相似
Semantic Segmentation:语义分割 ++
sensor:传感器
Stream Prefix:流前缀,指的是一个数据流中的前缀部分,即从流的起始位置开始一直到某个位置的所有数据。例如,如果一个数据流包含元素序列 {1, 2, 3, 4, 5},则该数据流的前缀可以是 {1}、{1, 2}、{1, 2, 3}、{1, 2, 3, 4} 或者 {1, 2, 3, 4, 5}。
sparsify:稀疏化,常用在机器学习中。
s-sparse numerical vector(s-稀疏数值向量):是指一个数值向量中只有少数几个非零元素,而其他元素都为零。在数学和信号处理中,稀疏性是指向量或矩阵中绝大部分元素都是零的特性。s-sparse numerical vector中的“s”表示非零元素的数量。稀疏性在很多领域中都有重要的应用,包括信号处理、压缩感知、统计学、机器学习等。具体来说,当数据具有稀疏性时,可以利用稀疏表示和处理方法来减少存储和计算成本,以及提高算法的效率
T
TECH REPORT(tech report):技术报告
Transformer:变形器
token:令牌,标记
- tokenize:令牌化
- tokenizer:令牌解析器
- Token Enbedding
transparency:透明度
V
vector:向量
validate:验证