I ran into the same problem, and thought I'd share a solution: multipart/form-data.

By sending a multipart form you send first as string your JSON meta-data, and then separately send as raw binary (image(s), wavs, etc) indexed by the Content-Disposition name.

Here's a nice tutorial on how to do this in obj-c, and here is a blog article that explains how to partition the string data with the form boundary, and separate it from the binary data.

The only change you really need to do is on the server side; you will have to capture your meta-data which should reference the POST'ed binary data appropriately (by using a Content-Disposition boundary).

Granted it requires additional work on the server side, but if you are sending many images or large images, this is worth it. Combine this with gzip compression if you want.

IMHO sending base64 encoded data is a hack; the RFC multipart/form-data was created for issues such as this: sending binary data in combination with text or meta-data.

The problem with UTF-8 is that it is not the most space efficient encoding. Also, some random binary byte sequences are invalid UTF-8 encoding. So you can't just interpret a random binary byte sequence as some UTF-8 data because it will be invalid UTF-8 encoding. The benefit of this constrain on the UTF-8 encoding is that it makes it robust and possible to locate multi byte chars start and end whatever byte we start looking at.

As a consequence, if encoding a byte value in the range [0..127] would need only one byte in UTF-8 encoding, encoding a byte value in the range [128..255] would require 2 bytes ! Worse than that. In JSON, control chars, " and \ are not allowed to appear in a string. So the binary data would require some transformation to be properly encoded.

Let see. If we assume uniformly distributed random byte values in our binary data then, on average, half of the bytes would be encoded in one bytes and the other half in two bytes. The UTF-8 encoded binary data would have 150% of the initial size.

Base64 encoding grows only to 133% of the initial size. So Base64 encoding is more efficient.

What about using another Base encoding ? In UTF-8, encoding the 128 ASCII values is the most space efficient. In 8 bits you can store 7 bits. So if we cut the binary data in 7 bit chunks to store them in each byte of an UTF-8 encoded string, the encoded data would grow only to 114% of the initial size. Better than Base64. Unfortunately we can't use this easy trick because JSON doesn't allow some ASCII chars. The 33 control characters of ASCII ( [0..31] and 127) and the " and \ must be excluded. This leaves us only 128-35 = 93 chars.

So in theory we could define a Base93 encoding which would grow the encoded size to 8/log2(93) = 8*log10(2)/log10(93) = 122%. But a Base93 encoding would not be as convenient as a Base64 encoding. Base64 requires to cut the input byte sequence in 6bit chunks for which simple bitwise operation works well. Beside 133% is not much more than 122%.

This is why I came independently to the common conclusion that Base64 is indeed the best choice to encode binary data in JSON. My answer presents a justification for it. I agree it isn't very attractive from the performance point of view, but consider also the benefit of using JSON with it's human readable string representation easy to manipulate in all programming languages.

If performance is critical than a pure binary encoding should be considered as replacement of JSON. But with JSON my conclusion is that Base64 is the best.

BSON (Binary JSON) may work for you. http://en.wikipedia.org/wiki/BSON

Edit: FYI the .NET library json.net supports reading and writing bson if you are looking for some C# server side love.

If you deal with bandwidth problems, try to compress data at the client side first, then base64-it.

Nice example of such magic is at http://jszip.stuartk.co.uk/ and more discussion to this topic is at JavaScript implementation of Gzip

yEnc might work for you:

http://en.wikipedia.org/wiki/Yenc

"yEnc is a binary-to-text encoding scheme for transferring binary files in [text]. It reduces the overhead over previous US-ASCII-based encoding methods by using an 8-bit Extended ASCII encoding method. yEnc's overhead is often (if each byte value appears approximately with the same frequency on average) as little as 1–2%, compared to 33%–40% overhead for 6-bit encoding methods like uuencode and Base64. ... By 2003 yEnc became the de facto standard encoding system for binary files on Usenet."

However, yEnc is an 8-bit encoding, so storing it in a JSON string has the same problems as storing the original binary data — doing it the naïve way means about a 100% expansion, which is worse than base64.

While it is true that base64 has ~33% expansion rate, it is not necessarily true that processing overhead is significantly more than this: it really depends on JSON library/toolkit you are using. Encoding and decoding are simple straight-forward operations, and they can even be optimized wrt character encoding (as JSON only supports UTF-8/16/32) -- base64 characters are always single-byte for JSON String entries. For example on Java platform there are libraries that can do the job rather efficiently, so that overhead is mostly due to expanded size.

I agree with two earlier answers:

• base64 is simple, commonly used standard, so it is unlikely to find something better specifically to use with JSON (base-85 is used by postscript etc; but benefits are at best marginal when you think about it)
• compression before encoding (and after decoding) may make lots of sense, depending on data you use

Just to add another option that we low level dinosaur programmers use...

An old school method that's been around since three years after the dawn of time would be the Intel HEX format. It was established in 1973 and the UNIX epoch started on January 1, 1970.

• Is it more efficient? No.
• Is it a well established standard? Yes.
• Is it human readable like JSON? Yes-ish and a lot more readable than most any binary solution.

The json would look like:

{
    "data": [
    ":10010000214601360121470136007EFE09D2190140",
    ":100110002146017E17C20001FF5F16002148011928",
    ":10012000194E79234623965778239EDA3F01B2CAA7",
    ":100130003F0156702B5E712B722B732146013421C7",
    ":00000001FF"
    ]
}

Smile format

It's very fast to encode, decode and compact

Speed comparison (java based but meaningful nevertheless): https://github.com/eishay/jvm-serializers/wiki/

Also it's an extension to JSON that allow you to skip base64 encoding for byte arrays

Smile encoded strings can be gzipped when space is critical

Since you're looking for the ability to shoehorn binary data into a strictly text-based and very limited format, I think Base64's overhead is minimal compared to the convenience you're expecting to maintain with JSON. If processing power and throughput is a concern, then you'd probably need to reconsider your file formats.

In depth

I dig a little bit more (during implementation of base128), and expose that when we send characters which ascii codes are bigger than 128 then browser (chrome) in fact send TWO characters (bytes) instead one :(. The reason is that JSON by defaul use utf8 characters for which characters with ascii codes above 127 are coded by two bytes what was mention by chmike answer. I made test in this way: type in chrome url bar chrome://net-export/ , select "Include raw bytes", start capturing, send POST requests (using snippet at the bottom), stop capturing and save json file with raw requests data. Then we look inside that json file:

• We can find our base64 request by finding string 4142434445464748494a4b4c4d4e this is hex coding of ABCDEFGHIJKLMN and we will see that "byte_count": 639 for it.
• We can find our above127 request by finding string C2BCC2BDC380C381C382C383C384C385C386C387C388C389C38AC38B this are request-hex utf8 codes of characters ¼½ÀÁÂÃÄÅÆÇÈÉÊË (however the ascii hex codes of this characters are c1c2c3c4c5c6c7c8c9cacbcccdce). The "byte_count": 703 so it is 64bytes longer than base64 request because characters with ascii codes above 127 are code by 2 bytes in request :(

So in fact we don't have profit with sending characters with codes >127 :( . For base64 strings we not observe such negative behaviour (probably for base85 too - I don check it) - however may be some solution for this problem will be sending data in binary part of POST multipart/form-data described in Ælex answer (however usually in this case we don't need to use any base coding at all...).

The alternative approach may rely on mapping two bytes data portion into one valid utf8 character by code it using something like base65280 / base65k but probably it would be less effective than base64 due to utf8 specification ...

function postBase64() {
  let formData = new FormData();
  let req = new XMLHttpRequest();

  formData.append("base64ch", "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/");
  req.open("POST", '/testBase64ch');
  req.send(formData);
}


function postAbove127() {
  let formData = new FormData();
  let req = new XMLHttpRequest();

  formData.append("above127", "¼½ÀÁÂÃÄÅÆÇÈÉÊËÌÍÎÏÐÑÒÓÔÕÖ×ØÙÚÛÜÝÞßàáâãäåæçèéêëìíîïðñòóôõö÷øùúûüý");
  req.open("POST", '/testAbove127');
  req.send(formData);
}

<button onclick=postBase64()>POST base64 chars</button>
<button onclick=postAbove127()>POST chars with codes>127</button>

(Edit 7 years later: Google Gears is gone. Ignore this answer.)

The Google Gears team ran into the lack-of-binary-data-types problem and has attempted to address it:

Blob API

JavaScript has a built-in data type for text strings, but nothing for binary data. The Blob object attempts to address this limitation.

Maybe you can weave that in somehow.

Just to add the resource and complexity standpoint to the discussion. Since doing PUT/POST and PATCH for storing new resources and altering them, one should remember that the content transfer is an exact representation of the content that is stored and that is received by issuing a GET operation.

A multi-part message is often used as a savior but for simplicity reason and for more complex tasks, I prefer the idea of giving the content as a whole. It is self-explaining and it is simple.

And yes JSON is something crippling but in the end JSON itself is verbose. And the overhead of mapping to BASE64 is a way to small.

Using Multi-Part messages correctly one has to either dismantle the object to send, use a property path as the parameter name for automatic combination or will need to create another protocol/format to just express the payload.

Also liking the BSON approach, this is not that widely and easily supported as one would like it to be.

Basically, we just miss something here but embedding binary data as base64 is well established and way to go unless you really have identified the need to do the real binary transfer (which is hardly often the case).

I suggest to use a non-standard JSON to save binary data. As we don't really see JSON, I don't care how it represent, thus all binary in a string is acceptable, the only character need to be escaped is just double-quote(") and backslash(\) itself, all other characters are acceptable, even \n do not necessarily need to be escaped, just keep binary 0d and 0a there is ok, it won't break JSON string as it is not " so it won't finish a string.

Your code can deal with all these binary data without problem, just care about " and \ is ok.

In Node.js, you can convert a Buffer to a string and back without any change:

const serialized = buffer.toString("binary")
const deserialized = Buffer.from(serialized, "binary")

If you want more reliability by sacrificing size, replace "binary" with "base64"

Data type really concerns. I have tested different scenarios on sending the payload from a RESTful resource. For encoding I have used Base64(Apache) and for compression GZIP(java.utils.zip.*).The payload contains information about film,an image and an audio file. I have compressed and encoded the image and audio files which drastically degraded the performance. Encoding before compression turned out well. Image and audio content were sent as encoded and compressed bytes [] .

Refer: http://snia.org/sites/default/files/Multi-part%20MIME%20Extension%20v1.0g.pdf

It describes a way to transfer binary data between a CDMI client and server using 'CDMI content type' operations without requiring base64 conversion of the binary data.

If you can use 'Non-CDMI content type' operation, it is ideal to transfer 'data' to/from a object. Metadata can then later be added/retrieved to/from the object as a subsequent 'CDMI content type' operation.

One other, more novel idea, is to encode the data via uuencode. It's a mostly deprecated, but it could still be an alternative. (Although perhaps not a serious one.)

While looking for a more space-efficient binary-to-text encoding I came across Base122. I’ve not played with it in detail but it seems to be a viable alternative to Base64.